Category Archives: Data modeling

Computational Classics: Finding errors in annotated ancient Greek texts with association rules mining

Published by:

This blog describes some experiments with ruleminer for finding morphological patterns in annotated data of ancient Greek texts. Ruleminer is a Python package for association rules mining, a rules-based machine learning method for discovering patterns in large data sets. The regex and dataframe approach in ruleminer (set out in this article) is used to enable a controlled search in the data set. Previously, I have used ruleminer mainly for quantitative data, but it might be worth investigating whether it is applicable to annotated (NLP) text data.

Finding annotation errors

The idea is to extract morphological patterns from annotated data and with these patterns detect annotation errors made by the NLP parser that was used for the annotations. Morphological patterns are recurrent relations between word forms and features, such as part of speech, tense, mood, case, number and gender. These recurrent relations can be expressed as association rules and mining algorithms can be used to find these relations. By looking at those situation where patterns were not satisfied it could be possible to identify errors made by the NLP parser. This is useful because many NLP pipelines use these annotation for subsequent analyses.

For many languages NLP parsers are available to annotate documents and determine lemmas and the morphological features of word forms within these documents. The performance of these models is often measured in the percentage of correct annotations against predefined treebanks, text corpora with annotations verified by linguists. Normally these models use deep learning algorithms and no model is yet able to achieve fully correct annotations; for many models these percentages lie around 95%. The Perseus model in Stanford Stanza for ancient Greek texts provides annotations with the following scores: universal part-of-speech tags (92,4%), treebank-specific part-of-speech tags (85,0%), treebank-specific morphological features (91,0%), and lemmas (88.3%).

Preprocessing steps

As a basis for ancient Greek data I took a number of dialogues of Plato (Apology, Crito, Euthydemos, Euthypron, Gorgias, Laws, Phaedon, Phaedrus, Republic, Symposium and Timaois). The text documents were annotated with the Perseus model (the model was originally not trained with this data) and the result was converted to the NLP Interchange Format (NIF 2.0) in RDF with OLiA annotations (all done with the nafigator package using the Stanford Stanza pipeline).

The data set consists of 312.955 annotated word forms. In order to apply ruleminer all words were extracted from the RDF-graph and stored in a Pandas DataFrame. Each word is a row in the DataFrame with the original text of the word (nif:anchorOf), the lemma of the word derived from the Perseus model (nif:lemma) and, for each feature (57 in total), whether or not the word is annotated with this feature (columns start with olia:, for example olia:Adverb).

The only changes to the original text were the deletion of diacritic signs acute and grave (like ά and ὰ) because sometimes the place of these signs changes when suffixes are added or deleted, which makes it harder to find patterns. All other diacritic signs were unchanged.

This notebook contains all examples mentioned below (and many more examples for a range of nominal forms).

Deriving morphological patterns in ancient Greek

In what follows I will show some examples of rules that can be found in the annotated text. To check whether word forms have the annotated features and whether they have multiple meanings (with different features) I used Perseus Greek Word Study Tool.

Particle patterns

Particles in ancient Greek are short word forms that are never inflected, i.e. the form or ending does not change. So ideal for finding morphological patterns. First we look at the morphological features. Let’s run the following expression to identify the OLiA annotations of the word γαρ:

if (({"nif:anchorOf"}=="γαρ")) then ({"olia:.*"}==1))

The olia:.* in this expression means all columns that start with olia:, i.e. OLiA annotations except the lemma and the original text. Ruleminer will match all these columns and evaluate the metrics of the resulting candidate rule. If the candidate rule satisfies predefined constraints (here a minimum confidence of 90% was used) it will be added to the resulting rules:

rule_definitionabs supportabs exceptionsconfidence
if ({“nif:anchorOf”}==”γαρ”) then ({“olia:Adverb”}==1)2165420.98097
Results of the expression: if (({“nif:anchorOf”}==”γαρ”)) then ({“olia:.*”}==1))

This produces one rule which states that the word γαρ is identified correctly as an adverb (the Perseus model maps particles to adverbs or conjunctions). This rule has a confidence of just over 98% (in 2.165 cases the word γαρ is annotated as an adverb). There are 42 exceptions, meaning that the word was not annotated as an adverb. These exceptions might point to situations where a word has different features (and meanings) depending on the context. In this case it is strange because γαρ has only one other meaning: the noun γάρος, which does not occur in Plato’s work. I therefore expect that these are all annotation errors by the Perseus model. We can check this by looking at the associations between word forms and their lemmas with the following rule.

if (({"nif:anchorOf"}=="γαρ")) then ({"nif:lemma"}==".*"))

Here are the results:

rule_definitionabs supportabs exceptionsconfidence
if ({“nif:anchorOf”}==”γαρ”) then ({“nif:lemma”}==”γαρ”)2187200.990938
if ({“nif:anchorOf”}==”γαρ”) then ({“nif:lemma”}==”γιγνομαι”)921980.004078
if ({“nif:anchorOf”}==”γαρ”) then ({“nif:lemma”}==”ἐγώ”)622010.002719
if ({“nif:anchorOf”}==”γαρ”) then ({“nif:lemma”}==”γιρ”)422030.001812
if ({“nif:anchorOf”}==”γαρ”) then ({“nif:lemma”}==”γιπτω”)122060.000453
Results of the expression: if (({“nif:anchorOf”}==”γαρ”)) then ({“nif:lemma”}==”.*”))

Here the first rule in the table is the only one that is correct. The others have very low confidence and are obvious errors by the Perseus model: γιρ and γιπτω are nonexistent words, and ἐγώ and γιγνομαι are not the lemmas of γαρ. So these are incorrect annotations, and errors by the NLP parser.

Next we consider the particles οὐ, οὐκ, οὐχ (negating particles) and μη (a particle indicating privation). In this case for each word two annotations are found, the olia:Adverb and the olia:Negation.

rule_definitionabs supportabs exceptionsconfidence
if ({“nif:anchorOf”}==”μη”) then ({“olia:Adverb”}==1)1753590.967439
if ({“nif:anchorOf”}==”μη”) then ({“olia:Negation”}==1)16941180.934879
if ({“nif:anchorOf”}==”οὐ”) then ({“olia:Adverb”}==1)133301.0000
if ({“nif:anchorOf”}==”οὐ”) then ({“olia:Negation”}==1)133210.9993
if ({“nif:anchorOf”}==”οὐκ”) then ({“olia:Adverb”}==1)124801.0000
if ({“nif:anchorOf”}==”οὐκ”) then ({“olia:Negation”}==1)124801.0000
if ({“nif:anchorOf”}==”οὐχ”) then ({“olia:Adverb”}==1)32201.0000
if ({“nif:anchorOf”}==”οὐχ”) then ({“olia:Negation”}==1)32201.0000
Results of the expression with negating particles

Most of the times the word μη is annotated as an adverb and as a negation. There are however a number of exceptions. Looking into this a bit further shows that the word is sometimes annotated as a subordinating conjunction and sometimes the lemma is mistakenly set to μεμω or εἰμι resulting in incorrect verb related annotations. Here are the lemmas in case the word is not an adverb:

rule_definitionabs supportabs exceptionsconfidence
if ({“nif:anchorOf”}==”μη”) then (({“olia:Adverb”}!=1)&({“nif:lemma”}==”μη”))4217700.0232
if ({“nif:anchorOf”}==”μη”) then (({“olia:Adverb”}!=1)&({“nif:lemma”}==”μεμω”)) 1617960.0088
if ({“nif:anchorOf”}==”μη”) then (({“olia:Adverb”}!=1)&({“nif:lemma”}==”εἰμι”))118110.0006

Pronoun patterns

The word form of Ancient Greek pronouns depend on the case and grammatical number. In most of the cases the personal pronoun does not have other meanings depending on the context, so this should lead to strong patterns. To run ruleminer with a list of expressions we can use the following code.

# personal pronouns
# first person, second person
pronouns = [
    'ἐγω', 'ἐμοῦ', 'ἐμοι', 'ἐμε', 
    'μου', 'μοι', 'με',
    'συ', 'σοῦ', 'σοι', 'σε', 
    'ἡμεῖς', 'ἡμῶν', 'ἡμῖν', 'ἡμᾶς',
    'ὑμεῖς', 'ὑμῶν', 'ὑμῖν', 'ὑμᾶς'
expressions = [
    'if (({"nif:anchorOf"}=="'+pn+'")) then ({"olia:Pronoun"}==1)'
    for pn in pronouns

Then we get, sorted with highest support, the following result

rule_definitionabs supportabs exceptionsconfidence
if ({“nif:anchorOf”}==”ἡμῖν”) then ({“olia:Pronoun”}==1)67101.0000
if ({“nif:anchorOf”}==”μοι”) then ({“olia:Pronoun”}==1)49940.9920
if ({“nif:anchorOf”}==”ἐγω”) then ({“olia:Pronoun”}==1)383360.9141
if ({“nif:anchorOf”}==”σοι”) then ({“olia:Pronoun”}==1)344110.9690
if ({“nif:anchorOf”}==”ἡμῶν”) then ({“olia:Pronoun”}==1)23001.0000
if ({“nif:anchorOf”}==”συ”) then ({“olia:Pronoun”}==1)229270.8945
if ({“nif:anchorOf”}==”ἡμᾶς”) then ({“olia:Pronoun”}==1)20801.0000
Results of pronoun patterns

Again this points to many errors in the Perseus model. Word forms like ἐγω, μοι and συ cannot be taken as anything other than a pronoun. However, the word σοι could also be a possessive adjective depending on the context.

Verbs patterns

We now have seen some easy examples with straightforward rules. For verbs we need more complex rules, but this is still feasible with ruleminer.

In ancient Greek if a verb is thematic, in present tense, indicative mood, plural and third person then the ending of that verb (if it is not contracted) is stem+ουσι(ν). To formulate a rule for this we want to keep the stem of the verb that was found in the antecedent (the if-part) and use it later on in the consequent of the rule (the then-part). This can be done by defining a regex group (by using with parentheses) in the following way:

if (({"nif:anchorOf"}=="(\w+[^εαο])ουσιν?")) then (({"nif:lemma"}=="\1ω"))

The if-part of the rule is true if the column nif:anchorOf matches the regex (\w+[^εαο])ουσιν?. The first part of this regex (between parenthesis) consists of one or more characters not ending with ε, α, and ο. This is the stem of the word and it is stored as a regex group (to be used in the consequent of the rule). The second part is ουσιν?, which is regex for either ουσι or ουσιν. The then-part of the rule is true is nif:lemma contains the stem of the word (in regex this is \1) plus ω.

The first five lines of the results (85 rules in total).

rule_definitionabs supportabs exceptionsconfidence
First five lines of the expression: if (({“nif:anchorOf”}==”(\w+[^εαο])ουσιν?”)) then (({“nif:lemma”}==”\1ω”))

Aggregate text analysis

Let’s end these examples with an aggregate analysis of the data set of all word forms with lemmas and morphological features. To find out if there is a prevalence in the text with respect to certain morphological features let’s run the following simple rules:

if (({"olia:ProperNoun"}==1)) then ({"olia:Neuter"}==1)
if (({"olia:ProperNoun"}==1)) then ({"olia:Feminine"}==1)
if (({"olia:ProperNoun"}==1)) then ({"olia:Masculine"}==1)

These rules identify the grammatical gender of all the proper nouns in the text (word forms that start with a capital letter and name people, places, things, and ideas). Here are the results:

rule_definitionabs supportconfidence

Almost 80% of the proper nouns in the text have masculine gender, and just over 10% have feminine gender. Remember that this is derived from Plato’s dialogues, so no surprise there. Most protagonists in the dialogues, if not all, are male and related word forms are therefore masculine. I specifically looked at the feminine proper nouns with more that five occurrences: they are geographical locations like Δῆλος (most frequent, 48 times), Συράκουσαι (7 times), Αίγυπτος (6 times). It also appeared that a number of male protagonists were incorrectly given annotations with feminine gender (Σιμμιας, 10 times and Μελητος, 6 times). Furthermore some word forms were mistakenly taken as pronoun, and that some pronouns did not have an annotation for gender (that is why it does not sum up).


As you can see this all works quite well. If a word form has one meaning then it is fairly easy to create reliable patterns and find erroneous annotations from a NLP parser. The main problem that cannot be solved in this approach (by looking at word forms only) is that in ancient Greek a word form can have more than one meaning, and therefore different morphological features, depending on the specific context of the word form. For example the meaning of a word form also depends on the (features of) preceding and following words in the sentence. To take that into account a different approach for mining is necessary.

I wonder whether it is feasible to automatically correct the output of the NLP parser in case of high confidence or humanly verified morphological patterns. That would increase the accuracy of the annotations. Furthermore, if the association rules are used for prediction then it might perhaps even be possible to construct a complete rules-based annotation model, and thereby replacing the deep learning model with a transparent rules-based approach.

So it must first be possible to create more complex rules that take into account the context of the word forms. This could be achieved by querying the RDF-graph directly to mine for reliable triple associations and with that find erroneous triples and missing triples in the graph. To be continued.

Multilingual termbases with metadata from reporting templates

Published by:

Domain-specific termbases are of great importance to many domain-specific NLP-tasks. They enable identification and annotation of terms in documents in situations where often not enough text is available to use statistical approaches. And more importantly, they form a step towards extracting structured facts from unstructured text data.

This blog shows how to construct and use multilingual termbases to annotate text from supervisory documents in different European languages with references to relevant parts of (quantitative) supervisory templates. By linking qualitative data (text) to quantitative data (numbers) we connect initially unstructured text data to data that are often to a high degree structured and well-defined in data point models and taxonomies. I will do this by constructing termbases that contain terminology data combined with linguistic data and metadata from the supervisory templates from different financial sectors.

For terminology data I will start with the IATE-database. Most terminology that is used in European quantitative reporting templates is based on and derived from European legislation. Having multilingualism as one of its founding principles is, the EU publishes terminology in the IATE-database in all official European languages to provide consistent translation of terms in European legislation. The IATE-database is published in the form of a file in TBX-format (TermBase eXchange). But termbases can also be in form of SKOS (Simple Knowledge Organization System, built upon the RDF-format). Both formats are data models that contain descriptions and properties of concepts and are to some extent interchangeable (see for example here).

For metadata on reporting templates I will use relevant XBRL Taxonomies in RDF-format (see here). Normally, XBRL Taxonomy are developed specifically for a single sector and therefore covers to some extent the financial terminology used within that sector. XBRL Taxonomies contain metadata of all data point in the reporting templates. From a XBRL Taxonomy a Data Point Model can be derived (that is: the taxonomy contains all definitions) and is often published together with the taxonomy which is only computer readable.

For linguistic data I will use the Python NLP package of Stanford Stanza, which provide pretrained NLP-models for all official European languages (in order of becoming an official EU language: Dutch, French, German, Italian (1958), Danish, English (1973), Greek (1981), Portuguese, Spanish (1986), Finnish, Swedish (1995), Czech, Estonian, Hungarian, Latvian, Lithuanian, Maltese, Polish, Slovak, Slovenian (2004), Bulgarian, Irish, Romanian (2007) and Croatian (2013)).

So we add semantic and linguistic structure to a terminology database. The resulting data structure is sometimes called an ontology, a taxonomy or a vocabulary, but these terms have no clear distinctive definitions. And moreover, the XBRL-people use the term taxonomy to refer to a structure that contains concepts with properties for definitions, labels, calculations and (table) presentations. To some extent it contains structured metadata of data points (i.e. the semantics of the data). Because of that you can say that it corresponds to an ontology within a Linked Data context. On the other hand a taxonomy within a Linked Data context (and everywhere else if I might add) is basically a description of concepts with sub-class relationships (concepts with hierarchical information). In the remainder of this blog I will use the term termbase for the resulting structure with semantic, linguistic and terminological data combined.

Constructing a termbase from IATE and XBRL

In a previous blog I have described how to set up a terminology database (termbase) specifically for insurance-related terms. Now I will add links from the concepts and terms in the IATE-database to the data point model of insurance reporting templates (thereby adding semantics to the termbase), and secondly I will add linguistic information at term-level like lemmas and part-of-speech tags to allow for easy usage in NLP-tasks. The TBX-format in which the IATE-database is published allows for storing references as well as linguistic data on term-level, so we can construct the termbase as a standalone file in TBX-format (another solution would be to add the terminology and linguistic information to the XBRL Taxonomy and use that as a basis).

The IATE-database currently contains almost 930.000 concepts and many of them have verbal expressions in multiple languages resulting in over 8.000.000 terms. A single (English) expression of a concept in the IATE-database looks like this.

<conceptEntry id="3539858">
  <descrip type="subjectField">insurance</descrip>
  <langSec xml:lang="en">
      <term>basic own funds</term>
      <termNote type="termType">fullForm</termNote>
      <descrip type="reliabilityCode">9</descrip>

Adding labels from the XBRL Taxonomy

For the termbase, we add every element in the XBRL Taxonomy that has a label (tables, concepts, elements, dimensions and members) to the termbase and we add an external cross reference to the template and the location in that template where the element is used (the row or column within the template). The TBX-format allows for fields called externalCrossReference which refer to a resource that is external to the terminology database. Then you get concept entries like this:

<conceptEntry id="">
  <descrip type="xbrlTypes">element</descrip>
  <xref type="externalCrossReference">S.,R0020</xref>
  <langSec xml:lang="en">
      <term>Basic own funds</term>
      <termNote type="termType">fullForm</termNote>
      <termNote type="termType">shortForm</termNote>

This means that the termbase concept with the id “…/s.” has English labels “Basic own funds” (full form) and “R0020” (short form). These are the labels of row 0020 of template S., i.e. in the template definition you will see Basic own funds on row 0020.

The template and rc-codes where elements refer to were extracted using SPARQL-queries on the XBRl Taxonomy in RDF-format.

Adding references between IATE- and XBRL-concepts

Now that we have added terms for labelled elements from the XBRL Taxonomy, the next step is to add cross references between the IATE-concepts and the XBRL-concepts. In the TBX-format the crossReference is a pointer to another related location, such as another entry or another term, in this case to a XBRL-concept. Below the references to the XBRL-concepts are added.

<conceptEntry id="3539858">
  <descrip type="subjectField">insurance</descrip>
  <langSec xml:lang="en">
      <term>basic own funds</term>
      <termNote type="termType">fullForm</termNote>
      <descrip type="reliabilityCode">9</descrip>
  <ref type="crossReference"></descrip>
  <ref type="crossReference"></descrip>
  <ref type="crossReference"></descrip>
  <ref type="crossReference"></descrip>

The IATE-concept with id 3539858 points to the domain item el#x7 (a domain in XBRL is a set of related items, in this case own funds items), and furthermore to (table) elements s., s. and s. These all refer to a single row or a single column within a template and the last one is given as an example above. It is the row in the table S. with label ‘Basic own funds’. IATE-concepts and XBRL-concepts are considered equal when their lowercase lemmas are the same (except for abbreviations).

For XBRL Taxonomy of Solvency 2 we find in this way 740 unique terms in the IATE-database that are identical to a XBRL-concept (in English), and 1.500 terms that occur in the labels of XBRL-concepts but are not identical, for example ‘net best estimate’ is a part of the label ‘Net Best Estimate of Premium Provisions’. For now I focused on identifying identical terms. How to process long labels with more than one term is a remaining challenge (this describes probably a useful approach).

Adding part-of-speech tags and lemmas

To make the termbase applicable for NLP-tasks we need to add additional linguistic information, such as part-of-speech patterns and lemmas. This improves the annotation process later on. Premium as an adjective has another meaning than premium as a common noun. By matching the PoS-pattern we get better matches (i.e. less false positives). And also the lemmas of a term will find terms irrespective of whether they are in grammatical singular or plural form. This is the concept to which PoS-patterns and lemma are added:

<conceptEntry id="3539858">
  <descrip type="subjectField">insurance</descrip>
  <langSec xml:lang="en">
      <term>basic own funds</term>
      <termNote type="termType">fullForm</termNote>
      <descrip type="reliabilityCode">9</descrip>
      <termNote type="termLemma">basic own fund</termNote>
      <termNote type="partOfSpeech">adj, adj, noun</termNote>
  <langSec xml:lang="it">
      <term>fondi propri di base</term>
      <termNote type="termType">fullForm</termNote>
      <descrip type="reliabilityCode">9</descrip>
      <termNote type="termLemma">fondo proprio di base</termNote>
      <termNote type="partOfSpeech">noun, det, adp, noun</termNote>

Annotating text in different languages

Because the termbase contains multilingual terms with references to templates and locations we are now able to annotate terms in documents in different European languages. Below you find some examples of what you can do with the termbase. I took an identical text from the Solvency 2 Delegated Acts in English, Finnish and Italian (the first part of article 52), converted the text to NAF and added annotations with the termbase (Nafigator has a function that processes a complete termbase and adds annotations to the NAF-file). This results in the following (using a visualizer from the spaCy package):

In English

Article 52 Mortality riskTerm S.,R0100 stress The mortality riskTerm S.,R0100 stress referred to in Article 77b(1)(f) of Directive 2009/138/EC shall be the more adverse of the 1. following two scenarios in terms of its impact on basic own fundsTerm S.,R0020: (a) an instantaneous permanent increase of 15 % in the mortality rates used for the calculation of the best estimateTerm S.,R0540; (b) an instantaneous increase of 0.15 percentage points in the mortality rates (expressed as percentages) which are used in the calculation of technical provisionsTerm S.,R0010 to reflect the mortality experience in the following 12 months.

In Finnish

52 artikla KuolevuusriskiinTerm S.,R0100 liittyvä stressi Direktiivin 2009/138/EY 77 b artiklan 1 kohdan f alakohdassa tarkoitetun, kuolevuusriskiinTerm S.,R0100 liittyvän stressin on 1. oltava seuraavista kahdesta skenaariosta se, jonka epäsuotuisa vaikutus omaan perusvarallisuuteenTerm S.,R0020 on suurempi: a) välitön, pysyvä 15 %:n nousu parhaan estimaatinTerm S.,R0540 laskennassa käytetyssä kuolevuudessa; b) välitön 0,15 prosenttiyksikön nousu prosentteina ilmaistussa kuolevuudessa, jota käytetään vakuutusteknisen vastuuvelanTerm S.,R0010 laskennassa ilmentämään havaittua kuolevuutta seuraavien 12 kuukauden aikana. Sovellettaessa 1 kohtaa kuolevuuden nousua sovelletaan ainoastaan vakuutuksiin, joissa kuolevuuden nousu johtaa

In German

Artikel 52 SterblichkeitsrisikostressTerm S.,R0100 Der in Artikel 77b Absatz 1 Buchstabe f der Richtlinie 2009/138/EG genannte SterblichkeitsrisikostressTerm S.,R0100 ist das im 1. Hinblick auf die Auswirkungen auf die BasiseigenmittelTerm S.,R0020 ungünstigere der beiden folgenden Szenarien: (a) plötzlicher dauerhafter Anstieg der bei der Berechnung des besten Schätzwerts zugrunde gelegten Sterblichkeitsraten um 15 %; (b) plötzlicher Anstieg der bei der Berechnung der versicherungstechnischen RückstellungenTerm S.,R0010 zugrunde gelegten Sterblichkeitsraten (ausgedrückt als Prozentsätze) um 0,15 Prozentpunkte, um die Sterblichkeit in den folgenden zwölf Monaten widerzuspiegeln.

In French

Article 52 Choc de risque de mortalitéTerm S.,R0100 Le choc de risque de mortalitéTerm S.,R0100 visé à l’article 77 ter, paragraphe 1, point f), de la directive 2009/138/CE 1. correspond au plus défavorable des deux scénarios suivants en termes d’impact sur les fonds propres de baseTerm S.,R0020: (a) une hausse permanente soudaine de 15 % des taux de mortalité utilisés pour le calcul de la meilleure estimationTerm S.,R0540; (b) une hausse soudaine de 0,15 point de pourcentage des taux de mortalité (exprimés en pourcentage) qui sont utilisés dans le calcul des provisions techniquesTerm S.,R0010 pour refléter l’évolution de la mortalité au cours des 12 mois à venir.

In Swedish

Artikel 52 DödsfallsriskstressTerm S.,R0100 Den dödsfallsriskstressTerm S.,R0100 som avses i artikel 77b.1 f i direktiv 2009/138/EG ska vara det som är mest negativt av 1. följande två scenarier i fråga om dess påverkan på kapitalbasen: (a) En omedelbar permanent ökning på 15 % av dödligheten som används för beräkning av bästa skattningenTerm S.,R0540. (b) En omedelbar ökning på 0,15 % av dödstalen (uttryckta i procent) som används i beräkningen av försäkringstekniska avsättningarTerm S.,R0010 för att återspegla dödligheten under de följande tolv månaderna.

In Italian

Articolo 52 Stress legato al rischio di mortalitàTerm S.,R0100 Lo stress legato al rischio di mortalitàTerm S.,R0100 di cui all’articolo 77 ter, paragrafo 1, lettera f), della direttiva 2009/138/CE è 1. il più sfavorevole dei due seguenti scenari in termini di impatto sui fondi propri di baseTerm S.,R0020: (a) un incremento permanente istantaneo del 15 % dei tassi di mortalità utilizzati per il calcolo della migliore stima; (b) un incremento istantaneo di 0,15 punti percentuali dei tassi di mortalità (espressi in percentuale) utilizzati nel calcolo delle riserve tecnicheTerm S.,R0010 per tener conto dei dati tratti dall’esperienza relativi alla mortalità nei 12 mesi successivi. Ai fini del paragrafo 1, l’incremento dei tassi di mortalità si applica soltanto alle polizze di assicurazione per le 2. quali tale incremento comporta un aumento delle riserve tecnicheTerm S.,R0010 tenendo conto di tutto quanto segue:

In the Italian text one reference is missing: the IATE-database does not yet contain an Italian translation for the English term best estimate. This happens because the IATE-database is far from complete. Not all terms are available in all languages, and the IATE-database probably does not contain all terminology from the reporting templates. And although the IATE-database is constantly updated, it might be necessary for certain use cases to add additional translations and concepts to the database.

This works for every XBRL Taxonomy, although every taxonomy has its own peculiarities (it’s a “standard” as one of my IT-colleagues likes to say). Following the procedure described above, I made the following termbases for insurance undertakings, credit institutions and Dutch pension funds based on the taxonomy versions mentioned:

  • Termbase of EIOPA Solvency 2, taxonomy version 2.6.0
  • Termbase of EBA CRD IV, taxonomy version
  • Termbase of DNB FTK, taxonomy version 2.3.0

These can be found on

Natural Language Processing in RDF graphs

Published by:

This blog shows how to store text data in a RDF graph and retrieve and analyze information from that graph. Resource Description Format (RDF) graphs are very suitable structures for storing Natural Language Processing (NLP) data. They enable combining NLP data with other data sets in RDF (such as legal entities data from the Global LEI Foundation and the EIOPA register of European insurance undertakings, terminology data, for example Solvency 2 terminology and data from XBRL reports); and they allow adding text semantics in the form of linguistic annotations, which enables NLP analyses simply by executing database queries.

Here is what I did. To get a proper amount of text data I web-scraped the entire website of De Nederlandsche Bank (text in webpages and in PDF documents, including speeches, press releases, research publications, sector information, dnbulletins, and all blogs by Maarten Gelderman and Olaf Sleijpen, consisting of over 4.000 documents). Text extraction from the web pages was done with the Python package newspaper3k (a great tip from my NLP colleagues from the Authority for Consumers and Markets). Text data was then converted to the NLP Annotation Format (NAF), for which I defined a RDF representation (implemented in the Nafigator package) to upload the data in a RDF triple-store. For the triple-store I used Ontotext’s GraphDB, one of the best RDF database currently available. Then, information can be retrieved from the graph database with SPARQL queries for all kinds of NLP analyses.

Using a triple-store for NLP data leads to an efficient retrieval process of text data, especially if you compare that to a process where you search through different annotation files. Triple-stores for RDF (and the new RDF-star) have become efficient and powerful solutions with the equal capabilities as property graphs but with advantages of RDF and ontologies.

I will describe two parts of this process that are not straightforward in detail: the RDF representation of NAF, and retrieving data from the graph database.

The NLP Annotation Format in RDF

The NLP Annotation Format is an easy format for storing text annotations (see here for links to the description). All documents that were scraped from the website were processed with the Python package Nafigator, that is able to convert PDF document and HTML-files to XML-files satisfying the NLP Annotation Format. Standard annotation layers with the raw text, word forms, terms, named entities and dependencies were added using the Stanford Stanza NLP processor.

In this representation every annotation (word forms, terms, named entities, etc.) of every document must have an Uniform Resource Identifier (URI). To do this, I used a prefix doc_xxx for each document in the document set. This prefix can, for example, be set by

@prefix doc_001: <> .

Which in this case is an identifier based on the domain of this blog. For web-scraped documents you might also use the original URL of the document. Furthermore, for the RDF representation of NAF a provisional RDF Schema with prefix naf-base was made with the basic properties en classes of NAF.

The basic structure is set out below. All examples provided below are derived from the file example.pdf in the Nafigator package (the first sentences of the first page starts with: ‘The Nafigator package … ‘).

Document and header

Every document has a header and pages.

doc_001:doc a naf-base:document ;
    naf-base:hasHeader doc_001:nafHeader ;
    naf-base:hasPages ( doc_001:page1 ) .

Here naf-base:document is a RDF Class and naf-base:hasHeader and naf-base:hasPages are RDF Properties. The three lines above state that doc_001:doc is a document with header doc_001:nafHeader and a single page doc_001:page1.

In the header all metadata of the document is stored, including all linguistics processors and models that were used in processing the document. Below you see the metadata of the NAF text layer and the document metadata.

doc_001:nafHeader a naf-base:header ;
    naf-base:hasLinguisticProcessors [ 
        naf-base:hasLayer naf-base:text ;
        naf-base:lp [ 
            naf-base:hasBeginTimestamp "2022-04-10T13:45:43UTC" ;
            naf-base:hasEndTimestamp "2022-04-10T13:45:44UTC" ;
            naf-base:hasHostname "desktop-computer" ;
            naf-base:hasModel "stanza_resources\\en\\tokenize\\" ;
            naf-base:hasName "text" ;
            naf-base:hasVersion "stanza_version-1.2.2" 
    ] ;
    naf-base:hasPublic [ 
        dc:format "application/pdf" ;
        dc:uri "data/example.pdf" 
    ] .

Sentences, paragraphs and pages

Here is an example of a sentence object with properties.

doc_001:sent1 a naf-base:sentence ;
    naf-base:isPartOf doc_001:para1, doc_001:page1 ;
    naf-base:hasSpan ( doc_001:wf1 doc_001:wf2 ...  doc_001:wf29 ) .

These three lines describe the properties of the RDF subject doc_001:sent1. The doc_001:sent1 identifies the RDF subject for the first sentence of the first document; the first line says that the subject doc_001:sent1 is a (rdf:type) sentence. The second line says that this sentence is a part of the first paragraph and the first page of the document. The span of the sentence contains a ordered list of word forms of the sentence: doc_001:wf1, doc_001:wf2 and so on.

Paragraphs and pages to which the sentences refer are defined in a similar way.

Word forms and terms

Of each word form the properties text, length and offset are defined. The word form is a part of a term, sentence, paragraph and page, and that is also defined for every word form. Take for example the the word form doc_001:wf2 defined as:

doc_001:wf2 a naf-base:wordform ;
    naf-base:hasText "Nafigator"^^rdf:XMLLiteral ;
    naf-base:hasLength "9"^^xsd:integer ;
    naf-base:hasOffset "4"^^xsd:integer ;
    naf-base:isPartOf doc_001:page1, 

In the next layer the terms of the word forms are defined, with their linguistic properties (lemma, grammatical number, part-of-speech and if applicable other properties such as verb voice and verb form). The term that refers to the word form above is

doc_001:term2 a naf-base:term ;
    naf-base:hasLemma "Nafigator"^^rdf:XMLLiteral ;
    naf-base:hasNumber olia:Singular ;
    naf-base:hasPos olia:ProperNoun ;
    naf-base:hasSpan ( doc_001:wf2 ) .

For the linguistic properties the OLiA ontology is used, which stand for Ontologies of Linguistic Annotations, an OWL taxonomy of data categories for linguistic annotations. The ontology contains precise definitions and interrelation between the linguistic categories. In this case the grammatical number (olia:Singular) and the part-of-speech tag (olia:ProperNoun) is included in the properties of this term. Depending of the term other properties are defined, for example verb forms. The span of the term refers back to the word forms (if you create a NAF ontology then you would define this as a transitive relationship, but for now, by including both relations we speed up the retrieval process).

Named entities

Next are the named entities that are stored in another NAF layer and here as separate subjects in the triple-store. An entity refers back to a term and has a certain type (organization, person, product, law, date and so on). The text of the entity is already stored in the term object so there is not need to include it here. External references could be added here, for example references to legal entities from Global LEI Foundation. Here is the example referring to the triples above.

doc_001:entity1 a naf-base:entity ;
    naf-base:hasType naf-entity:product ;
    naf-base:hasSpan ( doc_001:term2 ) .


Powerful NLP models exist that are able to derive relationships between words in within sentences. The dependencies are defined on the level of terms and stored in the dependency layer of NAF. In this RDF representation the dependencies are simply added to the terms.

doc_001:term3 a naf-base:term ;
    naf-rfunc:compound doc_001:term2 ;
    naf-rfunc:det doc_001:term1 .

The second and third line say that term3 (‘package’) forms a compound term with term2 (‘Nafigator’) and has its determinant in term1 (‘The’).

There are more annotation layers in NAF, but these are the most basic ones and if you have these, then many powerful NLP analyses already can be done.

Information retrieval from the RDF graph database

The conversion of text to RDF described above was applied to all webpages and documents of the website of DNB, 4.065 documents in total with 401.832 sentences containing 9.789.818 words. This text data led to over 221 million RDF triples in the tripe-store. I used a local database that was queried via a SPARQL endpoint. These numbers mentioned here can easily be extracted with SPARQL queries, for example to count the number of sentences we can use the SPARQL query:

SELECT (COUNT(?s) AS ?count) WHERE { ?s a naf-base:sentence . }

With this query all RDF subjects (the variable ?s) that are a sentence are counted and the result is stored in the variable ‘count’. The same can be done with other RDF subjects like word forms and documents.

The RDF representation described above allows you to store the content and annotations of a set of documents with their metadata in one single graph. You can then retrieve information from that graph from different perspectives and for different purposes.

Information retrieval

Suppose we want to find all references on the website with relations between ‘DNB’ and the verb ‘supervise’ by looking for sentences where ‘DNB’ is the nominal subject and ‘supervise’ is the lemma of the verb in the sentence. This is done with the following query

SELECT ?text
    ?term naf-base:hasLemma "supervise" .
	?term naf-rfunc:nsubj [naf-base:hasLemma "DNB" ] .
    ?term naf-base:hasSpan [ rdf:first ?wf ] .
    ?wf naf-base:isPartOf [ a naf-base:sentence ; naf-base:hasText ?text ].

It’s almost readable 🙂 The first line in the WHERE statement retrieves words that have ‘supervise’ as a lemma (this includes past, present and future tense and different verb forms). The second line narrows the selection down to where the nominal subject of the verb is ‘DNB’ (the lemma of the subject to be precise). The last two lines select the text of the sentences that includes the words that were found.

Execution of this query is done in a few milliseconds (on a desktop computer with a local database, nothing fancy) and results in 22 sentences, such as “DNB supervises adequate management of sustainability risks by financial institutions.”, “DNB supervises the cash payment system by providing information and guidance on the rules and procedures, data collection and examining compliance with the rules.”, and so on.

Term extraction

Terms are often multi-words and can be retrieved by part-of-speech tags and dependencies. Suppose we want to retrieve all two-words terms of the form adjective, common noun. Part-of-speech tags are defined in the terms layer. In the graph also the relation between the terms is defined, in this case by an adjectival modifier (amod) relation (the common noun is modified by an adjective). Then we can define a query that looks for exactly that: two words, an adjective and a common noun, where the mutual relationship is of an adjectival modifier. This is expressed in the first three lines in the WHERE-clause below. The last two lines retrieve the text of the words.

SELECT DISTINCT ?w1 ?w2 (count(*) as ?c)
    ?term1 naf-base:hasPos olia:CommonNoun .
    ?term2 naf-base:hasPos olia:Adjective .
    ?term1 naf-rfunc:amod ?term2 .
    ?term1 naf-base:hasSpan [ rdf:first/naf-base:hasText ?w1 ] .
    ?term2 naf-base:hasSpan [ rdf:first/naf-base:hasText ?w2 ] .
} GROUP BY ?w1 ?w2

Note that in the query a count of the number of occurrences of the term in the output and sort the output according to this count has been added.

Most often the term ‘monetary policy’ was found (2.348 times), followed by ‘financial institutions’ (1.734 times) and ‘financiële instellingen’ (Dutch translation of financial institution, 1.519 times), and so on. In total more than 127.000 of these patterns were found on the website (this is a more complicated query and took around 10 seconds). In this way all kinds of term patterns can be found, which can be collected in a termbase (terminology database).

Opinion extraction

I will give here a very simple example of opinion extraction based on part-of-speech tags. Suppose you want to extract sentences that contain the authors (or someone else’s) subjective opinion. You can look a the grammatical subject and the verb in a sentence, but you can also look at whether a sentence contains something like ‘too high’ or ‘too volatile’ (which often indicates a subjective content). In that case we have the word ‘too’ (an adverb) followed by an adjective, with mutual relation of adverbial modifier (advmod). In the Dutch language this has exactly the same form. The following query extracts these sentences.

SELECT ?text
    ?term1 naf-base:hasPos olia:Adjective .
    ?term2 naf-base:hasSpan [ rdf:first/naf-base:hasText "too" ] .
    ?term1 naf-rfunc:advmod ?term2 .
    ?term1 naf-base:hasSpan [ rdf:first ?wf1 ] .
    ?sent1 naf-base:hasSpan [ rdf:rest*/rdf:first ?wf1 ] .
    ?sent1 a naf-base:sentence .
    ?sent1 naf-base:hasText ?text .

With the last three lines the text of the sentence that includes the term is found (the output of the query). With the documents of the website of DNB, the output contains sentences like: “It is also clear that CO2 emissions are still too cheap and must be priced higher to sufficiently curtail emissions” and “Firms end up being too large” (in total 343 sentences in 0.3 seconds).

The examples shown here are just for illustrative purposes and do not always lead to accurate results, but they show that information extraction can be done fairly easy (if you know SPARQL) and reasonably quick. Once the data is stored into a graph database, named entities can be matched with other internal or external data sources and lemmas of terms can be matched with concept-based terminology databases. Then you have a graph where the text is not only available on a simple string-level but also, and more importantly, on a conceptual level.

UPDATE: I have written a follow-up on this blog here.

The Solvency termbase for NLP

Published by:

This blog describes a way to construct a terminology database for the insurance supervision knowledge domain. The goal of this termbase is provide a reliable basis to extract insurance supervision terminology within different NLP analyses.

The terminology of solvency and insurance supervision forms an expert domain of terminology based on economics, mathematics, accounting and finance terminologies. Like probably many other knowledge domains the terminology used is very technical and specific. Some terms are only used within this domain with only a limited number of occurrences (which often hinders the use of statistical methods for finding terms). And many other words have general meanings outside the domain that do not coincide with the specific meanings within the domain. Translation of terms from this specific domain often requires extensive knowledge about the meaning and use of these terms.

What is a termbase?

A termbase is a database containing terminology and related information. It consists of concepts with their verbal designations (terms, i.e. single words or composed of multi-word strings) of a specific knowledge domain, often in different languages. It contains the full form of concepts, but also abbreviations, synonyms and variants and additional information of concepts, such as definitions and external references. To indicate the accuracy or completeness often a reliability code is added to individual terms of a concept. A proper termbase is an important terminology tool to achieve standardization of information and consistent use of (translations) of concepts in documents. And because of that, they are often used by professional translators.

The European Union translates legal documents in all member state languages and uses for this one common publicly available termbase: the IATE (Interactive Terminology for Europe) terminology database. The IATE termbase is used in the EU institutions and agencies since 2004 for the collection, dissemination and management of EU-specific terminology. This helps to avoid divergences in the application of European Law within Europe (there exists a vast amount of literature on the effects on language ambiguity in European legislation). The translations of European legislation are therefore of the highest quality with strong consistency between different directives, regulations, delegated and implementing acts and so on. This termbase is very useful for information extraction from documents and for linking terminology concepts between different documents. They can be extended with abbreviations, synonyms and common variants of terms.

Termbases is very useful for information extraction from documents and for linking terminology concepts between different documents. They can be extended with abbreviations, synonyms and common variants of terms.

The Solvency termbase for NLP

To create a first Solvency termbase for NLP purposes, I extracted terms from Solvency 2 Delegated Acts in a number of languages, looked up these terms in the IATE database and copied the corresponding concepts. It often happens that for one language the same term refers to different concepts (for example, the term ‘balance’ means something different in chemistry and in accounting). But if for one legal document the terms from different languages refer to the same concept, then we probably have the right concept (that was used in the translation of the legal document). So, the more references from the same legal document, the more reliable the term-concept relation is. And if we have the proper term-concept relationship, we automatically have all reliable translations of that concept.

Term extraction was done with part-of-speech patterns (such as adj-noun and adj-noun-noun patterns). To do this, for every language the Delegated Acts was converted to the NLP Annotation Format (NAF). The functionality for conversion to NAF and for extracting terms based on pos patterns is part of the nafigator package. As an NLP engine for nafigator, I used the Stanford Stanza package that contains tokenizers and part-of-speech models for every European language. The termbase itself was made with the terminator repository (currently under construction).

For terms in Dutch, I also added to the termbase additional part-of-speech tags, lemma’s and morphological properties from the Lassy Klein-corpus from the Instituut voor de Nederlandse taal (Dutch Language Institute). This data set consists of approximately 1 million words with manually verified syntactic annotations. I expanded this data set with solvency related words. Linguistical properties of terms of other languages can be added it a reliable data set is available.

Below, you see one concept from the resulting termbase (the concept of which ‘solvency capital requirement’ is the English term) in TermBase eXchange format (TBX). This is an international standard (ISO 30042:2019) for the representation of structured concept-oriented terminological data, based on xml.

<conceptEntry id="249">
 <descrip type="subjectField">insurance</descrip>
 <langSec xml:lang="nl">
   <termNote type="partOfSpeech">noun</termNote>
   <note>source: ../naf-data/data/legislation/Solvency II Delegated Acts - NL.txt (#hits=331)</note>
   <termNote type="termType">fullForm</termNote>
   <descrip type="reliabilityCode">9</descrip>
   <termNote type="lemma">solvabiliteits_kapitaalvereiste</termNote>
   <termNote type="grammaticalNumber">singular</termNote>
    <termNote type="component">solvabiliteits-</termNote>
    <termNote type="component">kapitaal-</termNote>
    <termNote type="component">vereiste</termNote>
 <langSec xml:lang="en">
   <termNote type="termType">abbreviation</termNote>
   <descrip type="reliabilityCode">9</descrip>
   <term>solvency capital requirement</term>
   <termNote type="termType">fullForm</termNote>
   <descrip type="reliabilityCode">9</descrip>
   <termNote type="partOfSpeech">noun, noun, noun</termNote>
   <note>source: ../naf-data/data/legislation/Solvency II Delegated Acts - EN.txt (#hits=266)</note>
 <langSec xml:lang="fr">
   <term>capital de solvabilité requis</term>
   <termNote type="termType">fullForm</termNote>
   <descrip type="reliabilityCode">9</descrip>
   <termNote type="partOfSpeech">noun, adp, noun, adj</termNote>
   <note>source: ../naf-data/data/legislation/Solvency II Delegated Acts - FR.txt (#hits=198)</note>
   <termNote type="termType">abbreviation</termNote>
   <descrip type="reliabilityCode">9</descrip>

You see that the concept contains a link to the IATE database entry with the definition of the concept (the link in this blog actually works so you can try it out). Then a number of language sections contain terms of this concept for different languages. The English section contains the term SCR as an English abbreviation of this concept (the French section contains the abbreviation CSR for the same concept). For every term the part-of-speech tags were added (which are not part of the IATE database) and, for Dutch only, with the lemma and grammatical number of the term and its word components. These additional linguistical attributes allow easier use within NLP analyses. Furthermore, as a note the number of all occurrences in the original legal document are included.

The concept entry contains related terms in all European languages. In Greek the SCR is κεφαλαιακή απαίτηση φερεγγυότητας, in Irish it is ‘ceanglas maidir le caipiteal sócmhainneachta’ (although the Solvency 2 Delegated Acts is not available in the Irish language), in Portuguese it is ‘requisito de capital de solvência’, in Estonian ‘solventsuskapitalinõue’, and so on. These are reliable translations as they are used in legal documents of that language.

The termbase contains all terms from the Solvency 2 Delegated Acts that can be found in the IATE database. In addition, terms that were not found in that database are added with the termNote “NewTerm”, to indicate that this term has yet to be reviewed by a knowledge domain expert. This would also be the way to add synonyms and variants of terms.

The Solvency termbase basically allows to scan for a Solvency 2 concept in a document in any of the 23 European languages (given that it is in the IATE database). This is of course an initial approach to construct a termbase to test whether it is feasible and practical. The terminology that insurance undertakings use in their solvency reports is very likely to differ from the one used in legal documents. I will be testing this with a number of documents to identify Solvency 2 terminology to get an idea of how many synonyms and variants are missing.

Besides this Solvency termbase, it is in the same way possible to construct a Climate termbase based on the European Climate Law (a European regulation from 2021). This law contains a large number of climate-related terminology and is available in all European languages. A Climate termbase gives the possibility to extract climate-related information from all kinds of documents. Furthermore, we have the Sustainable Finance Disclosure Regulation (a European regulation also from 2021) for environmental, social, and governance (ESG) terminology, which could provide a starting point for an ESG termbase. And of course I eagerly await the European Regulation on Artificial Intelligence.

Converting XBRL to RDF-star

Published by:

Lately I have been working on the conversion of XBRL instances and related taxonomy schemas and linkbases to RDF and RDF-star. In these semantic data formats, you can link data in XBRL data with other data sources and you can query the data in a fairly easy manner. RDF-star is an extension of RDF that in some situations allows a more compact description of linked data, and by that it narrows the gap between RDF and property graphs. How this works, I will show in this blog using the XBRL taxonomy definitions as an example.

In a previous blog I showed that XBRL instance facts can be converted to RDF and visualized as a network. The same can be done with the related taxonomy elements. An XBRL taxonomy consists of concepts and relations between concepts that define calculations, presentations, labels and definitions. The concepts are laid down (mostly) in XML schemas and the relations in linkbases using XML schemas and XLinks. By converting the XBRL taxonomy to RDF, the XBRL fact data is linked to its corresponding metadata in the taxonomy.


There has been done some work on the conversion of XBRL to RDF, most notably by Dave Raggett. His project xbrlimport, written in C++ and available on SourceForge, converts XBRL data to RDF triples. His approach is clean and straightforward and reuses the original namespaces of the XBRL data (with some obvious elements translated to predicates with RDF namespaces).

I used Raggett’s xbrlimport as a starting point, translated it to Python, added XBRL items that were introduced after publication of the code and improved a number of things. The code is now for example able to convert all EIOPA’s Solvency 2 taxonomy elements with all metadata available to RDF format. This code is available under the same license as xbrlimport (GNU General Public License) as a Python Package on You can take an XBRL instance with corresponding taxonomy (in the form of a zip-file) and convert the contents to RDF and RDF-star. This code will look up any references (URIs) in the XBRL instance to the taxonomy in the zip-file and convert the relevant files to RDF.

Let’s look at some examples of the Solvency 2 taxonomy converted to RDF. The RDF triple of an arbitrary XBRL concept from the Solvency 2 taxonomy looks like this (in turtle format):

    rdf:type xbrli:monetaryItemType ;
    xbrli:periodType """instant"""^^rdf:XMLLiteral ;
    model:creationDate """2014-07-07"""^^xsd:dateTime ;
    xbrli:substitutionGroup xbrli:item ;
    xbrli:nillable "true"^^xsd:boolean ;

This example describes the triples of concept s2md_met:mi362 (a Solvency 2 metric). With these triples we have exactly the same data as in the related XML file but now in the form of triples. Namespaces are derived from the XML file (except rdf:type) and datatypes are transformed to RDF datatypes with proper RDF syntax.

This can be done with all concepts used to which the facts of an XBRL instance refer. If you have facts in RDF format, then in RDF these concept are automatically linked with the concepts in the taxonomy because the URIs of the concepts are the same. This creates a network of facts with all related metadata of the facts.

An XBRL taxonomy also contains links that relate concepts to each other for several purposes (to provide labels, definitions, presentations and calculations) . An example of a link is the following.

_:link2 arcrole:concept-label [
    xl:type xl:link ;
    xl:role role2:link ;
    xl:from s2md_met:mi362 ;
    xl:to s2md_met:label_s2md_mi362 ;
    ] .

The link relates concept mi362 with label mi362 by creating a new subject _:link2 with predicate arcrole:concept-label and an object which contains all data about the link (including the xl:from and xl:to and the attributes of the link). This way of introducing a new subject to specify a link between two concepts is called reification and a bit artificial because you would like to link the concept directly with the label, such as

s2md_met:mi281 arcrole:concept-label s2md_met:label_s2md_mi281

However, then you are unable in RDF to link the attributes (like the order and the role) to the predicates. It is one of the disadvantages of the current RDF format. There appears to be no easy way to do this in RDF, other than by using this artificial reification approach (some other solutions exist like the singleton property approach, but all of them have disadvantages.)

The new RDF-star format

Recently, the RDF-star working group published their first Draft Community Report. In this report they introduced new RDF-star and SPARQL-star specifications. These new specifications, although not yet a W3C standard, enable more compact specification of linked datasets and simpler graphs and less nodes.

Let’s look what this means for the XBRL linkbases with the following example. Suppose we have the following link definition.

_:link1 arcrole:breakdown-tree [
    xl:from _:s2md_a1 ;
    xl:to _:s2md_a1.root ;
    xl:type xl:link ;
    xl:role tab:S. ;
    xl:order "0"^^xsd:decimal ;
    ] .

The subject in this case is _:link1 with predicate arcrole:breakdown-tree, so this link describes a part of a table template. It points to a subject with all the information of the link, i.e. from, to, type, role and order from the xl namespace. Note that there is no triple with _:s2md_a1 (xl:from) as a subject and _:s2md_a1.root (xl:to) as an object. So if you want to know the relations of the concept _:s2md_a1 you need to look at the link triples and look for entries where xl:from equals the concept.

With the new RDF-star specifications you can just add the triple and then add properties to the triple as a whole, so the example would read

_:s2md_a1 arcrole:breakdown-tree _:s2md_a1.root .

<<_:s2md_a1 arcrole:breakdown-tree _:s2md_a1.root>> 
    xl:role tab:S. ;
    xl:order "0"^^xsd:decimal ;

Which is basically what we need to define. If you now want to know the relations of the subject _:s2md_a1 then you just look for triples with this subject. In the visual presentation of the RDF dataset you will see a direct link between the two concepts. This new RDF format also implies simplifications of the SPARQL queries.

This blog has become a bit technical but I hope you see that the RDF-star specification allows a much needed simplification of RDF triples. I showed that the conversion of XBRL taxonomies to RDF-star leads to a smaller amount of triples and also to less complex triples. The resulting taxonomy triples lead to less complex graphs and can be used to derive the XBRL labels, template structures, validation rules and definitions, just by using SPARQL queries.

Europe’s insurance register linked to the GLEIF RDF dataset

Published by:

Number 7 of my New Year’s Resolutions list reads “only use and provide linked data”. So, to start the year well, I decided to do a little experiment to combine insurance undertakings register data with publicly available legal entity data. In concrete terms, I wanted to provide the European insurance register published by EIOPA (containing all licensed insurance undertakings in Europe) as linked data with the Legal Entity data from the Global Legal Entity Identifier Foundation (GLEIF). In this blog I will describe what I have done to do so.

The GLEIF data

The GLEIF data consists of information on all legal entities in the world (entities with a Legal Entity Identifier). A LEI is required by any legal entity who is involved with financial transactions or operates within the financial system. If an organization needs a LEI then it requests one at a local registration agent. For the Netherlands these are the Authority for the Financial Markets (AFM), the Chamber of Commerce (KvK) and the tax authority and others. GLEIF receives data from these agents in each country and makes the collected LEI data available in a number of forms (api, csv, etc).

The really cool thing is that in 2019, together with, GLEIF developed an RDFS/OWL Ontology for Legal Entities, and began in 2020 to publish regularly the LEI data as a linked RDF dataset on (see, you need a (free) account to obtain the data). At the time of this writing, the size of the level 1 data (specifying who is who) is around 10.2 Gb with almost 92 million triples (subject-predicate-object), containing information about entity name, legal form, headquarter and legal address, geographical location, etc. Also related data such as who owns whom is published in this forms.

The EIOPA insurance register

The European Supervisory Authority EIOPA publishes the Register of Insurance undertakings based on information provided by the National Competent Authorities (NCAs). The NCA in each member state is responsible for authorization and registration of the insurance undertakings activities. EIOPA collects the data in the national registers and publishes an European insurance register, which includes more than 3.200 domestic insurance undertakings. The register contains entity data like international and commercial name, name of NCA, addresses, cross border status, registration dates etc. Every insurance undertaking requires a LEI and the LEI is included in the register; this enables us to link the data easily to the GLEIF data.

The EIOPA insurance register is available as CSV and Excel file, without formal naming and clear definitions of column names. Linking the register data with other sources is a tedious job, because it must be done by hand. Take for example the LEI data in the register, which is referred to with the column name ‘LEI’; this is perfectly understandable for humans, but for computers this is just a string of characters. Now that the GLEIF has published its ontologies there is a proper way to refer to a LEI, and that is with the Uniform Resource Identifier (URI), or in a short form gleif-L1:LEI.

The idea is to publish the data in the European insurance register in the same manner; as linked data in RDF format using, where applicable, the GLEIF ontology for legal entities and creating an EIOPA ontology for the data that is unique for the insurance register. This allows users of the data to incorporate the insurance register data into the GLEIF RDF dataset and thereby extending the data available on the legal entities with the data from the insurance register.

Creating triples from the EIOPA register

To convert the EIOPA insurance register to linked data in RDF format, I did the following:

  • extract from the GLEIF RDF level 1 dataset on all insurance undertakings and related data, based on the LEI in the EIOPA register;
  • create a provisional ontology with URIs based on (this should ideally be done by EIOPA, as they own the domain name in the URI);
  • transform, with this ontology, the data in the EIOPA register to triples (omitting all data from the EIOPA register that is already included in the GLEIF RDF dataset, like names and addresses);
  • publish the triples for the insurance register in RDF Turtle format on

Because I used the GLEIF ontology where applicable, the triples I created are automatically linked to the relevant data in the GLEIF dataset. Combining the insurance register dataset with the GLEIF RDF dataset results in a set where you have all the GLEIF level 1 data and all data in the EIOPA insurance register combined for all European insurance undertakings.

Querying the data

Let’s look what we have in this combined dataset. Querying the data in RDF is done with the SPARQL language. Here is an example to return the data on Achmea Schadeverzekeringen.

{ ?s gleif-L1:hasLegalName "Achmea schadeverzekeringen N.V." . 
  ?s ?p ?o .}

The query looks for triples where the predicate is gleif-base:hasLegalName and the object is Achmea Schadeverzekeringen N.V. and returns all data of the subject that satisfies this constraint. This returns (where I omitted the prefix of the objects):

    rdf#type                             LegalEntity
    gleif-base:hasLegalJurisdiction      NL  
    gleif-base:hasEntityStatus           EntityStatusActive  
    gleif-l1:hasLegalName                Achmea Schadeverzekeringen     
    gleif-l1:hasLegalForm                ELF-B5PM
    gleif-L1:hasHeadquartersAddress      L-72450067SU8C745IAV11-LAL  
    gleif-L1:hasLegalAddress             L-72450067SU8C745IAV11-LAL  
    gleif-base:hasRegistrationIdentifier BID-RA000463-08053410
    rdf#type                             InsuranceUndertaking
    eiopa-base:hasRegisterIdentifier     IURI-De-Nederlandsche-Bank-

We see that the rdf#type of this entity is LegalEntity (from the GLEIF data) and the jurisdiction is NL (this has a prefix that refers to the ISO 3166-1 country codes). The legal form refers to another subject called ELF-B5PM. The headquarters and legal address both refer to the same subject that contains the address data of this entity. Then there is a business identifier to the registration data. The last two lines are added by me: a triple to specify that this subject is not only a LegalEntity but also an InsuranceUndertaking (defined in the ontology), and a triple for the Insurance Undertaking Register Identifier (IURI) of this subject (also defined in the ontology).

Let’s look more closely at the references in this list. First the legal form of Achmea (i.e. the predicate and objects of legal form code ELF-B5PM). Included in the GLEIF data is the following (again omitting the prefix of the object):

rdf#type                         EntityLegalForm  
rdf#type                         EntityLegalFormIdentifier  
gleif-base:identifies            ELF-B5PM  
gleif-base:tag                   B5PM  
gleif-base:hasCoverageArea       NL  
gleif-base:hasNameTransliterated naamloze vennootschap  
gleif-base:hasNameLocal          naamloze vennootschap  
gleif-base:hasAbbreviationLocal  NV, N.V., n.v., nv

With the GLEIF data we have this data on all legal entity forms of insurance undertakings in Europe. The local abbreviations are particularly handy as they help us to link an entity’s name extracted from documents or other data sources with its corresponding LEI.

If we look more closely at the EIOPA Register Identifier IURI-De-Nederlandsche-Bank-W1686 then we find the register data of this Achmea entity:

owl:a                         InsuranceUndertakingRegisterIdentifier
gleif-base:identifies         L-72450067SU8C745IAV11
eiopa-base:hasNCA                          De Nederlandsche Bank  
eiopa-base:hasInsuranceUndertakingID       W1686  
eiopa-base:hasEUCountryWhereEntityOperates NL  
eiopa-base:hasCrossBorderStatus            DomesticUndertaking  
eiopa-base:hasRegistrationStartDate        23/12/1991 01:00:00  
eiopa-base:hasRegistrationEndDate          None  
eiopa-base:hasOperationStartDate           23/12/1991 01:00:00  
eiopa-base:hasOperationEndDate             None

The predicate gleif-base:identifies refers back to the subject where gleif-L1:hasLegalName equals the Achmea entity. The other predicates are based on the provisional ontology I made that contains the definitions of the attributes of the EIOPA insurance register. Here we see for example that W1686 is the identifier of this entity in DNB’s insurance register.

Let me give a tiny example of the advantage of using linked data. The GLEIF data contains the geographical location of all legal entities. With the combined dataset it is easy to obtain the location for the insurance undertakings in, for example, the Netherlands. This query returns entity names with latitude and longitude of the legal address of the entity.

SELECT DISTINCT ?name ?lat ?long
WHERE {?sub rdf:type eiopa-base:InsuranceUndertaking ;
            gleif-base:hasLegalJurisdiction CountryCodes:NL ;
            gleif-L1:hasLegalName ?name ;
            gleif-L1:hasLegalAddress/gleif-base:hasCity ?city .
       ?geo gleif-base:hasCity ?city ;
            geo:lat ?lat ; 
            geo:long ?long .}

This result can be plotted on a map, see the link below. If you click on one of the dots then the name of the insurance undertaking will appear.

All queries above and the code to make the map are included in the notebook EIOPA Register RDF datase – SPARQL queries.

The provisional ontology I created is not yet semantically correct and should be improved, for example by incorporating data on NCAs and providing formal definitions. And other data sources could be added, for example the level 2 dataset to identify insurance groups, and the ISIN to LEI relations that are published daily by GLEIF.

By introducing the RDFS/OWL ontologies, the Global LEI Foundation has set an example on how to publish (financial) entity data in an useful manner. The GLEIF RDF dataset reduces the time needed to link the data with other data sources significantly. I hope other organizations that publish financial entity data as part of their mandate will follow that example.

Converting supervisory reports to Semantic Webs: from XBRL to RDF

Published by:

A growing number of supervisory reports across Europe are based on the XML Extensible Business Reporting Language standard (XBRL). Financial entities such as banks, insurance undertakings and pension institutions are required to submit their reports to their supervisors in this format.

XBRL is a language for modeling, exchanging and automatically processing business and financial information. Reports in this format (called instance documents) are based on metadata (set out in taxonomies) that add semantic meaning to the data points that are reported. You can choose different implementations but overall an XBRL taxonomy provides a semantically rich data model and that has always been one of the main advantages of XBRL.

However, in its raw format (an XML file) each report is basically a machine readable document with a tree structure that does not enable easy integration with related data from other sources or integration with text documents and their contents.

In this blog, I will show that converting the XBRL reports to another format allows easier integration and understanding. That other format is based on Semantic Webs. It has been shown that XBRL converted to Semantic Webs can be done without any loss of information (see for example this article). So if we convert the XBRL format to a Semantic Web then we keep the structure and the meaning provided by the taxonomy. The result is basically a graph and this format enables integration with other linked data that is much easier.

A Semantic Web consists of formats and technologies that are rather old (from a computer science perspective): it originated around the same time as XBRL, some twenty years ago. And because it tried to solve similar problems (lack of semantic meaning in the World Wide Web) as the XBRL standard (lack of semantic meaning in business and financial data), to some extent it is based on similar concepts. It was however developed completely separate from XBRL.

The general concept of a Semantic Webs, where data is linked together to provide semantic meaning, is also known as a knowledge graph.

How does a Semantic Web work? One of the formats of the Semantic Web is the Resource Description Framework (RDF), originally designed as a metadata data model. RDF was adopted as a World Wide Web Consortium recommendation in 1999. The RDF 1.0 specification was published in 2004, and RDF 1.1 followed in 2014.

The RDF format is based on expressions in the form of subject-predicate-object, called triples. The subject and object denote (web) resources and the predicate denotes the relationship between the subject and the object. For example the expression ‘Spinoza has written the book Ethica Ordine Geometrico Demonstrata’ in RDF is a triple with a subject denoting “Spinoza”, a predicate denoting “has written”, and an object denoting “the book Ethica Ordine Geometrico Demonstrata”. This is a different approach than for example object-oriented models with an entity (Spinoza), attribute (book) and value (Ethica).

The RDF format could potentially solve some problems with the XBRL format. To explain this, I converted an XBRL-instance (a test instance file from EIOPA for Solvency 2) to RDF format.

Below you see the representation of one arbitrary data point in the report (called a fact) in RDF format and visualized as a network (I used the Python package networkx). The predicates contain the complete web resource so I limited the name to the last word to make it readable.

The red node is the starting point of the data point. The red labels on the lines describe the predicate between subject and object. You see that the fact (subject) ‘has decimals’ (predicate) 2 (object), and furthermore has unit EUR, has value 838522076.03, has type metmi503 (an internal code describing Payments for reported but not settled claims) and some other properties.

The data point also has a so-called context that defines the entity to which the fact applies, the period of time the fact is relevant (in this case 2019-12-31) and also a scenario, which consists of additional metadata of the data point. In this case we see that the data point is related to statutory accounts, non-life and health non-STL, direct business and accepted during the period (and a node without a label).

All facts in every XBRL instance are structured in this way, which means that for example you can search all facts with the label statutory accounts. Furthermore, because XBRL uses namespaces you can unambiguously identify predicates and objects in the report. For example, you see that the entity node has an identifier (starting with 0LFF1…) and a scheme (17442). The scheme refers to the web resource for the ISO standard 17442 which specifies the Legal Entity Identifier (LEI), so the entity is unambiguously identified with the given (LEI-)code. If you add other XBRL instances with references to that entity then the data is automatically linked because other instances will contain exactly the same entity node.

The RDF representation of the XBRL fact above is:

_:provenance1 xl:instance "filename".
_:unit_u xbrli:measure iso4217:EUR.
  xl:provenance :provenance1;
  xl:type xbrli:fact; 
  rdf:type s2md_met:mi503;
  rdf:value "838522076.03"^^xsd:decimal;
  xbrli:decimals "2"^^xsd:integer;
  xbrli:unit :unit_u; 
  xbrli:context :context_BLx79_DIx5_IZx1_TBx28_VGx84.
  xl:type xbrli:context;
  xbrli:entity [
    xbrli:identifier "0LFF1WMNTWG5PTIYYI38";
  xbrli:scenario [
    xbrldi:explicitMember "s2c_LB:x79"^^rdf:XMLLiteral;
    xbrldi:explicitMember "s2c_DI:x5"^^rdf:XMLLiteral;
    xbrldi:explicitMember "s2c_RT:x1"^^rdf:XMLLiteral;
    xbrldi:explicitMember "s2c_LB:x28"^^rdf:XMLLiteral;
    xbrldi:explicitMember "s2c_AM:x84"^^rdf:XMLLiteral;
  xbrli:instant "2019-12-31"^^xsd:date.

Instead of storing the data in separate templates with often unclear code names you can also convert the XBRL data to one large Semantic Web where all facts are linked together. The RDF format thus provides a graph model which allows easier integration and visualization (and, for me at least, easier understanding). It allows adding and linking data from other sources, such as Solvency 2 documents and external data, in the same graph.

Typically, supervisory reports consists of thousands of data points and supervisors receive reports from many entities each period. How would you store that information? I think that the natural way to store an XBRL instance is not a relational database but a graph database (like graphDB or Neo4j). These databases can store the facts with all the metadata in a structured way and enable to query the graph efficiently. Next blog, I will explore graph databases and query languages for XBRL reports converted to the RDF format.

Word2vec models for SFCRs

Published by:

Word2vec is a well-known algorithm for natural language processing that often leads to surprisingly good results, if trained properly. It consists of a shallow neural network that maps words to a n-dimensional number space, i.e. it produces vector representations of words (so-called word embeddings). Word2vec does this in a way that words used in the same context are embedded near to each other (their respective vectors are close to each other). In this blog I will show you some of the results of word2vec models trained with Wikipedia and insurance-related documents.

One of the nice properties of a word2vec model is that it allows us to do calculations with words. The distance between two word vectors provides a measure for linguistic or semantic similarity of the corresponding words. So if we calculate the nearest neighbors of the word vector then we find similar words of that word. It is also possible to calculate vector differences between two word vectors. For example, it appears that for word2vec model trained with a large data set, the vector difference between man and woman is roughly equal to the difference between king and queen, or in vector notation kingman + woman = queen. If you find this utterly strange then you are not alone. Besides some intuitive and informal explanations, it is not yet completely clear why word2vec models in general yield these results.

Word2vec models need to be trained with a large corpus of text data in order to achieve word embeddings that allow these kind of calculations. There are some large pre-trained word vectors available, such as the GloVe Twitter word vectors, trained with 2 billion tweets, and the word2vec based on google news (trained with 100 billion words). However, most of them are in the English language and are often trained on words that are generally used, and not domain specific.

So let’s see if we can train word2vec models specifically for non-English European languages and trained with specific insurance vocabulary. A way to do this is to train a word2vec model with Wikipedia pages of a specific language and additionally train the model with sentences we found in public documents of insurance undertakings (SFCRs) and in the insurance legislation. In doing so the word2vec model should be able to capture the specific language domain of insurance.

The Wikipedia word2vec model

Data dumps of all Wikimedia wikis, in the form of a XML-files, are provided here. I obtained the latest Wikipedia pages and articles of all official European languages (bg, es, cs, da, de, et, el, en, fr, hr, it, lv, lt, hu, mt, nl, pl pt, ro, sk, sl, fi, sv). These are compressed files and their size range from 8.6 MB (Maltese) to 16.9 GB (English). The Dutch file is around 1.5 GB. These files are bz2-compressed; the uncompressed Dutch file is about 5 times the compressed size and contains more than 2.5 million Wikipedia pages. This is too large to store into memory (at least on my computer), so you need to use Python generator-functions to process the files without the need to store them completely into memory.

The downloaded XML-files are parsed and page titles and texts are then processed with the nltk-package (stop words are deleted and sentences are tokenized and preprocessed). No n-grams were applied. For the word2vec model I used the implementation in the gensim-package.

Let’s look at some results of the resulting Wikipedia word2vec models. If we get the top ten nearest word vectors of the Dutch word for elephant then we get:

In []: model.wv.most_similar('olifant', topn = 10)
Out[]: [('olifanten', 0.704888105392456),
        ('neushoorn', 0.6430075168609619),
        ('tijger', 0.6399451494216919),
        ('luipaard', 0.6376790404319763),
        ('nijlpaard', 0.6358680725097656),
        ('kameel', 0.5886276960372925),
        ('neushoorns', 0.5880545377731323),
        ('ezel', 0.5879943370819092),
        ('giraf', 0.5807977914810181),
        ('struisvogel', 0.5724758505821228)]

These are all general Dutch names for (wild) animals. So, the Dutch word2vec model appears to map animal names in the same area of the vector space. The word2vec models of other languages appear to do the same, for example norsut (Finnish for elephant) has the following top ten similar words: krokotiilit, sarvikuonot, käärmeet, virtahevot, apinat, hylkeet, hyeenat, kilpikonnat, jänikset and merileijonat. Again, these are all names for animals (with a slight preference for Nordic sea animals).

In the Danish word2vec model, the top 10 most similar words for mads (in Danish a first name derived from matthew) are:

In []: model.wv.most_similar('mads', topn = 10)
Out[]: [('mikkel', 0.6680521965026855),
        ('nicolaj', 0.6564826965332031),
        ('kasper', 0.6114416122436523),
        ('mathias', 0.6102851033210754),
        ('rasmus', 0.6025335788726807),
        ('theis', 0.6013824343681335),
        ('rikke', 0.5957099199295044),
        ('janni', 0.5956574082374573),
        ('refslund', 0.5891965627670288),
        ('kristoffer', 0.5842193365097046)]

Almost all are first names except for Refslund, a former Danish chef whose first name was Mads. The Danish word2vec model appears to map first names in the same domain in the vector space, resulting is a high similarity between first names.

Re-training the Wikipedia Word2vec with SFCRs

The second step is to train the word2vec models with the insurance related text documents. Although the Wikipedia pages for many languages contain some pages on insurance and insurance undertakings, it is difficult to derive the specific language of this domain from these pages. For example the Dutch word for risk margin does not occur in the Dutch Wikipedia pages, and the same holds for many other technical terms. In addition to the Wikipedia pages, we should therefore train the model with insurance specific documents. For this I used the public Solvency and Financial Condition Reports (SFCRs) of Dutch insurance undertakings and the Dutch text of the Solvency II Delegated Acts (here is how to download and read it).

The SFCR sentences are processed in the same manner as the Wikipedia pages, although here I applied bi- and trigrams to be able to distinguish insurance terms rather than separate words (for example technical provisions is a bigram and treated as one word, technical_provisions).

Now the model is able to derive similar words to the Dutch word for risk margin.

In []: model.wv.most_similar('risicomarge')
Out[]: [('beste_schatting', 0.43119704723358154),
        ('technische_voorziening', 0.42812830209732056),
        ('technische_voorzieningen', 0.4108312726020813),
        ('inproduct', 0.409644216299057),
        ('heffingskorting', 0.4008549451828003),
        ('voorziening', 0.3887258470058441),
        ('best_estimate', 0.3886040449142456),
        ('contant_maken', 0.37772029638290405),
        ('optelling', 0.3554660379886627),
        ('brutowinst', 0.3554105758666992)]

This already looks nice. Closest to risk margin is the Dutch term beste_schatting (English: best estimate) and technische_voorziening(en) (English: technical provision, singular and plural). The relation to heffingskorting is strange here. Perhaps the word risk margin is not solely being used in insurance.

Let’s do another one. The acronym skv is the same as scr (solvency capital requirement) in English.

In []: model.wv.most_similar('skv')
Out[]: [('mkv', 0.6492390036582947),
        ('mcr_ratio', 0.4787723124027252),
        ('kapitaalseis', 0.46219778060913086),
        ('mcr', 0.440476655960083),
        ('bscr', 0.4224048852920532),
        ('scr_ratio', 0.41769397258758545),
        ('ðhail', 0.41652536392211914),
        ('solvency_capital', 0.4136047661304474),
        ('mcr_scr', 0.40923237800598145),
        ('solvabiliteits', 0.406883180141449)]

The SFCR documents were sufficient to derive an association between skv and mkv (English equivalent of mcr), and the English acronyms scr and mcr (apparently the Dutch documents sometimes use scr and mcr in the same context). Other similar words are kapitaalseis (English: capital requirement) and bscr. Because they learn from context, the word2vec models are able to learn words that are synonyms and sometimes antonyms (for example we say ‘today is a cold day’ and ‘today is a hot day’, where hot and cold are used in the same manner).

For an example of a vector calculation look at the following result.

In []: model.wv.most_similar(positive = ['dnb', 'duitsland'], 
                             negative = ['nederland'], topn = 5)
Out[]: [('bundesbank', 0.4988047778606415),
        ('bundestag', 0.4865422248840332),
        ('simplesearch', 0.452720582485199),
        ('deutsche', 0.437085896730423),
        ('bondsdag', 0.43249475955963135)]

This function finds the top five similar words of the vector DNBNederland + Duitsland. This expression basically asks for the German equivalent of De Nederlandsche Bank (DNB). The model generates the correct answer: the German analogy of DNB as a central bank is the Bundesbank. I think this is somehow incorporated in the Wikipedia pages, because the German equivalent of DNB as a insurance supervisor is not the Bundesbank but Bafin, and this was not picked up by the model. It is not perfect (the other words in the list are less related and for other countries this does not work as well). We need more documents to find more stable associations. But this to me is already pretty surprising.

There has been some research where the word vectors of word2vec models of two languages were mapped onto each other with a linear transformation (see for example Exploiting Similarities among Languages for Machine Translation, Mikolov, et al). In doing so, it was possible to obtain a model for machine translation. So perhaps it is possible for some European languages with a sufficiently large corpus of SFCRs to generate one large model that is to some extent language independent. To derive the translation matrices we could use the different translations of European legislative texts because in their nature these texts provide one of the most reliable translations available.

But that’s it for me for now. Word2vec is a versatile and powerful algorithm that can be used in numerous natural language applications. It is relatively easy to generate these models in other languages than the English language and it is possible to train these models that can deal with the specifics of insurance terminology, as I showed in this blog.

Pattern discovery in Solvency 2 data (2)

Published by:

Pattern discovery is a general technique to derive patterns from the data itself, without using explicit knowledge about the data. Previously, I described how to discover patterns in Solvency 2 data, for example that values of column x are most of the time or (almost) always equal to the values of column y, or higher, or lower, etc. I also showed how to find columns whose values are the sum of values of other columns, which often occurs in financial reports.

Here I will take pattern discovery one step further and discover constant ratios between values of columns. That is, patterns where the quotient between values of two columns equals a constant.

An example will clarify this. We take the own funds data of Dutch insurance undertakings (sheet 20 of the Excel file) and select the last year of data.

df = get_sheet(20)
df = df.xs(datetime(2017,12,31), axis = 0, level = 1, drop_level = False)

Now we run the generate-function from the insurlib package on Github. The generate-function allows the parameter pattern = “ratio”. Make sure the minimum support is higher than one otherwise you end up with a lot of ratio-patterns. Patterns are generated with the following code, with minimum support 3 and minimum confidence 10%.

rules = patterns.generate(dataframe = df,
                          pattern = "ratio",
                          parameters = {"min_confidence": 0.1, 
                                        "min_support": 3}))
rules = list(rules)

The first two patterns in the list are

['ratio', '9/20', ['mcr , total', 'scr , total']], 19, 120, 0.1367]


['ratio', '1/4', ['mcr , total', 'scr , total']], 36, 103, 0.259]

The two patterns state that the quotient between the values of the columns ‘mcr , total’ and ‘scr , total’ is 9/20 in 19 cases with confidence of almost 14% and is 1/4 (with 120 exception to this rule) is 36 cases with confidence of almost 26% (with 103 exceptions). In Solvency 2, the acronym MCR stands for Minimum Capital Requirement, and SCR stand for Solvency Capital Requirement (the solvency 2 risk-based capital requirement with VaR 99,5%). So the identified ratios are 9/20 (i.e. 45%) and 1/4 (i.e. 25%).

It is not desirable that the MCR is too low or too high in relation to the SCR. Solvency 2 legislation therefore prescribes that the MCR is bounded between 25% and 45% of the SCR. We see that these legal boundaries are found in the data itself. We have 36 insurance undertakings where the ratio is 25%, 19 undertaking where the ratio is 45%. The rest (should) have a ratio somewhere between 25% and 45%.

The algorithm applies a brute-force search for ratios between columns. It uses the fraction module of Python’s standard library to convert every quotient to a fraction, which enables the use of rational number arithmetic. For example

>>> Fraction(9,4)

The fraction expression you see in the pattern definition is the string representation of the fraction object.

A common problem with dealing with quotients is that often the original data set contains quotients like 0.2499999 and 0.2500001 because values are rounded off to euros. In the fraction module, this can be overcome by putting a limit on the denominator (default is 1e7, but you can pass this as a parameter). The result is that the nearest fraction is found given the limit on the denominator.

The other rules that were found in the own funds data are

['ratio', '9/100', ['available and eligible own funds|total eligible own funds to meet the mcr , tier 2', 'scr , total']], 3, 12, 0.2]

['ratio', '1/5', ['available and eligible own funds|total eligible own funds to meet the mcr , tier 2', 'mcr , total']], 12, 3, 0.8]

['ratio', '3/20', ['available and eligible own funds|total eligible own funds to meet the scr , tier 3', 'scr , total']], 6, 15, 0.2857]

['ratio', '1/3', ['available and eligible own funds|total eligible own funds to meet the scr , tier 3', 'mcr , total']], 3, 18, 0.1429]

These ratio-patterns are boundaries on the available and eligible own funds and are prescribed by Solvency 2 legislation. In the Netherlands, they have a relatively low support because a limited number of Dutch insurance undertakings hit the legal boundaries of tier two and tier three capital.

Here we looked at the own funds data. But many more patterns of this type can be found if you run the complete data set of Dutch insurance undertakings. It takes some time but then you will find in total 121 ratio-patterns (with minimum support of 3) covering loss-absorbing capacity of deferred taxes, impact of transitional measures, operational risk capital, etc. And of course other data sets than Solvency 2 data are possible.

These ratio-patterns work well because the Solvency 2 legislation contains legal boundaries that are represented in the data. Also other nonlegal constant ratios can be found, such as entity-specific ratios. You can find ratio-patterns for each insurance undertaking separately and then signal when a pattern is violated. This would work with a relatively low limit on the denominator of the fractions, because we want to find constant ratios with a high confidence. But for this to work well you will need more than two years of data.

A brute-force approach works well for the Dutch public Solvency 2 data set (around 1270 dimensions). But for high dimensional data it will take some time because of the possible combinations. Perhaps smart ways exist to detect ratios more easily, for example via statistical correlations.

The ratio-patterns that were discovered could be related to the Solvency 2 legislation by automatically reading ratios and percents in the text. And in the same manner, entity-specific ratios could be related to the SFCR documents of that entity. But that is for future work.

Pattern discovery in Solvency 2 data (1)

Published by:

This blog describes some results with algorithms for pattern discovery in Solvency 2 quantitative reports. The main idea here is to automatically uncover patterns that are present in these reports. We want to find patterns that represent widely or commonly occurring situations, and possibly represent business rules and relations prescribed by the underlying legislation. If we are able to find these patterns, then we are also able to identify when data satisfy and do not satisfy the patterns.

How can we find these patterns? Initially, I started with association rules. This is a rule-based machine learning approach that leads to transparent and explainable results. However, for this approach the quantitative data has to be encoded to a set of features (for example by replacing every quantitative value with an appropriate nominal value) and with high dimensional data this quickly becomes computationally expensive.

After some increments I decided to program a pattern discovery algorithm especially designed for analyzing the quantitative data points of reports. The goal was to speed up the process, while maintaining the association rules mining approach and performance measures. Below I will give some examples of how and which patterns can be found. The code is part of the insurlib package on Github. I used the public Solvency 2 data of Dutch insurers to find patterns in the quantitative data of these insurers (but applications to other data sets are possible).

First read the pandas and the insurlib package. The patterns-part consists of functions to generate patterns in numerical columns of dataframes.

import pandas as pd
from insurlib import patterns

Now read the public Solvency 2 data as was described in How to analyze public Solvency 2 data of Dutch insurers (by reading the Excel file and defining the read_sheet function). To recall, the Excel consists of the following worksheets:

  • Worksheet 14: balance sheet
  • Worksheet 15: premiums – life
  • Worksheet 16: premiums – non-life
  • Worksheet 17: technical provisions – life
  • Worksheet 18: technical provisions – non-life
  • Worksheet 19: transition and adjustments
  • Worksheet 20: own funds
  • Worksheet 21: solvency capital requirements – 1
  • Worksheet 22: solvency capital requirements – 2
  • Worksheet 23: minimum capital requirements
  • Worksheet 24: additional information life
  • Worksheet 25: additional information non-life
  • Worksheet 26: additional information reinsurance

Example 1: comparing two dataframes

Suppose we want to compare the worksheet balance sheet and the worksheet non-life technical provision and find the relations between the contents in these worksheets.

df1 = get_sheet(14)
df2 = get_sheet(18)
df2.columns = [str(df2.columns[i]) for i in range(len(df2.columns))]

The last line is to convert the multiple level columns to one level (so that we can compare it more easily with other dataframes).

You can generate patterns with the generate-function. Patterns that are found have the ‘association rule’-structure P -> Q. If you input two dataframes (P_dataframe and Q_dataframe) then all columns of P_dataframe are compared to all columns of Q_dataframe. The pattern we are looking for is ‘=’, so patterns with corresponding values are found. You can also use other patterns, such as ‘<‘, ‘<=’ , ‘>’, ‘>=’ and ‘!=’. Also a dict of parameters is used as input, with in this case the minimum confidence and the minimum support.

rules = patterns.generate(P_dataframe = df1,
                          Q_dataframe = df2,
                          pattern = "=",
                          parameters = {"min_confidence": 0.75, 
                                        "min_support": 10}))
rules = list(rules)
print("Number of rules: " + str(len(rules)))
Number of rules: 2

The output is a generator that we can convert to a list of rules. Two rules are found in this case. Let’s look a the first rule.

    'assets|reinsurance recoverables from:|non-life and health 
    similar to non-life , solvency ii value', 
    "('total non-life obligation', 'Technical provisions -
    total|Recoverable from reinsurance contract/SPV and Finite
    Re after the adjustment for expected losses due to counterparty
    - total')"], 

The first pattern states that the value of the reinsurance recoverables on the asset side for non-life on the balance sheet equates the value of the recoverable from reinsurance contracts in the technical provision for non-life obligations in the technical provisions sheet. The rule has confidence of almost 97%, and a support of 127. This means that there are 127 occurrences of this pattern in the data and in 97% of all occurrences (with nonzero data points) the patterns holds. We also see that in four cases the patterns are not present, i.e. the reinsurance recoverables does not equate to the recoverable in the technical provision (these are presumably data errors).

The second rule that was found reads:

    'liabilities|technical provisions – non-life , solvency ii value',
    "('total non-life obligation', 'Technical provisions -
    total|Technical provisions - total')"], 

This rule also has a high confidence. It says that the value of the technical provisions for non-life in the balance sheet equals the value of the total technical provisions in the technical provision sheet. This is a plain consistency rule between the sheets. The six exceptions are presumably data errors.

Both rules were, at the moment of publication, not part of the automatic and predefined validation rules of the Solvency II reports (otherwise the confidence would be 100%), as part of the XBRL-taxonomy. But by analyzing the reports in this manner we were able to uncover them automatically.

Example 2: patterns of sums

Often financial reports contain sums within the report. We can analyze the column names to detect potential sums (often a hierarchy in the columns name can be identified), but we can also find patterns of sums. The following code does that. We input the balance sheet dataframe and let the algorithm search for ‘sum’-patterns. The parameters sum_elements states the maximum elements in the sum (in this case three).

rules = patterns.generate(dataframe = df1,
                          pattern = "sum",
                          parameters = {"sum_elements": 3})
rules = list(rules)
print("Number of rules: " + str(len(rules)))
Number of rules: 7

Let’s take a look at the first rule:

    ['assets|investments (other than assets held for index-linked and 
     unit-linked contracts)|equities|equities - listed , solvency ii 
     'assets|investments (other than assets held for index-linked and 
     unit-linked contracts)|equities|equities - unlisted , solvency ii
    'assets|investments (other than assets held for index-linked and 
     unit-linked contracts)|equities , solvency ii value'], 

The rule states that the sum of the listed and unlisted equities equals to the equities (so equities are either listed or non-listed). This rule has a confidence of 100%, and there is presumably a validation rule within the reports. Six rules in this structure were found in this way. This is however somewhat computationally expensive.

Example 3: patterns with a given value

The last example searches for patterns with specific values. In this case we want to know in how many cases the investments are higher than zero. We can do this in the following way. We input the dataframe like in example 2 and we add a parameter columns and set it to the name of the column we want to investigate (in fact you can input a list of columns).

P = ['assets|investments (other than assets held for index-linked and 
      unit-linked contracts) , solvency ii value']
 rules = patterns.generate(dataframe = df1, 
                           pattern = ">", 
                           columns = P, 
                           value = 0, 
                           parameters = {'min_confidence': 0.75,
                                         'min_support': 1})
 rules = list(rules)
     'assets|investments (other than assets held for index-linked and 
     unit-linked contracts) , solvency ii value', 

The value of investment is, with confidence of 96%, higher than zero. In eleven cases the value is not higher than zero. This rule has a high confidence because, normally, insurers invest premiums collected for insurance policies into a wide range of investment assets. If no list of columns is added, patterns in all numerical columns in the dataframe returned.

The aim of these examples is to give a general idea of pattern discovery in Solvency 2 quantitative data. Numerous patterns can be found in this way by using the complete data set. And by using the measures confidence and support we can find patterns that are not exactly perfect but do provide information about the data, without taking recourse to statistical methods. Data errors and specific situations that lead to exceptions in the data are not expressions of pure randomness and should therefore not be analyzed by statistical methods. With these patterns we are able to reconstruct basic patterns in the data that provide information about the data.

Of course, many improvements are possible in order to find more complex patterns (and that why there is a (1) in the title of this blog). Presumably all existing validation rules can be found in this manner, and much more. Hopefully I will be able to implement these improvements and present them in a new blog.