Untangling the complexities of legal documentation is an imperative need for non practitioners of the legal profession. The terminology used in the domain is complex and it usually requires expert knowledge to be fully understood, since the legal framework is constantly being updated and the meaning of terms vary accordingly. Non-proprietary Automatic Terminology Extraction (ATE) tools are required in this particular domain in which documents contain private and sensitive data. This paper describes methods for obtaining accurate legal terms from labour law corpora, overcoming the di culties present in the area, and also analyses the peculiarities of the legal jargon, speci cally, in Spanish language. The performed experiments, executed with JATE, a wellknown open source library in the ATE literature, are still preliminary, but promising.
As evidenced by legal summaries published in Eurlex1, legal documentation such as laws, directives, decrees and even regular notices are often complex to understand by general public: legal terminology can be a real headache for people that are not used to this type of language. At the same time, many citizens and businesses across the European Union have to deal with signi cant compliance problems: breach of contracts, overdue debts, excessive working hours, etc.
With the aim of softening such complications, this paper proposes to retrieve legal terms from Spanish corpora through Automatic Term Extraction (ATE) techniques. Such legal terms are understood as words and multi-word expressions with a speci c meaning within a legal text; collections of such terms are considered terminologies. These terminologies could be afterwords interlinked with other language resources to share information, which contributes to obtain de nitions, translations and context, easing the comprehension of legal documentation.
ATE is a well-known technique in Natural Language Processing (NLP) that supports important tasks such as machine translation, speech recognition or information retrieval, to mention but a few. Several tools are already o ering this kind of technology; however, many of them still present unresolved limitations, such as noise generation, disambiguation issues or performance, delays with big corpora, which are some of the most frequent. Here, two additional limitations have been identi ed as crucial when dealing with legal information: domain speci city and data privacy. Current ATE tools nd di culties when extracting highly-speci c legal expressions and, on the other hand, they might include personal data such as proper names and identity numbers in the resulting term lists.
For this reason, this contribution proposes to con gure the tool JATE [
The rst use case has been developed within the Lynx project, which has provided the Spanish legal corpus2 used in the experiments. Lynx project3 is an H2020 Innovation Action towards the creation of a Knowledge Graph of legal and regulatory data from di erent jurisdictions and languages. This Legal Knowledge Graph interlinks multilingual legal information and provides explanations and context for legal expressions appearing in each document.
The terms extracted with the model developed here contribute to the creation of this platform that help European citizens understand legal documentation without having to invest time and money in speci c legal consulting.
JATE was originally developed for English terminology extraction; since this use case deals with Spanish corpora, it was also required an extension of the tool to cover Spanish language.
This paper is organised as follows: Section 3 presents the related work on term extraction technologies, Section 2 exposes the motivation behind this contribution and the analysis of the problem, Section 4 describes the experiments performed during this work and nally Section 5 includes conclusions and future work to be performed in next stages. 2
Some hints of the motivation behind this project have already been presented in the introduction: legal documentation is an unsolvable puzzle for laymen and non-practitioners of the legal domain.
In addition, many of the available legal language resources are not in machine readable formats yet. The major part is published as PDF, which hinders their look-up, and some of them are still distributed in physical format.
The work proposed here tackles this situation by researching on a most accurate terminology extraction methodology from those machine-readable legal
documents with the aim of creating new language resources that can be processed by the newest technologies.
A higher level of accuracy in the extracted terms means less time and money invested in human post processing of the resulting term list. However, the most important advantage lays on the e cacy of Semantic Web technologies: accurate legal terms can easily be linked with other legal language resources, o ering users more information such as translations, synonyms, context and related terms. This additional information can help them improve their comprehension of the legal documentation.
Available tools present several limitations, and only a few of them are open source. This means a major drawback since one of the ideas of this contribution is to avoid using proprietary software to address privacy issues.
Also, a common shortcoming of web based applications is personal data management. Legal documents contain private data and it may not be safe to upload them on a web-based tool. In the Lynx project, both public and sensitive documents are being handled, thus, data privacy is an important factor to keep in mind.
Since JATE is an open source framework, it allows the use of a given algorithm selected by the user and even the creation of new ones. This work includes the extension of the tool to cover terminology extraction from Spanish corpora, the analysis of the labour law corpora provided by Lynx partners to discover speci c patterns of legal language and the con guration of such patterns in the tool with the aim of extracting more accurate terms. 3
Prior to the availability of term extraction tools, this activity was carried out by domain experts and terminology professionals. Despite being the most accurate manner for terminology extraction, it is also the most expensive and timeconsuming. Taking into account the amount of information generated nowadays, human terminology extraction is unpractical.
For this reason, automatic term extraction technologies have been extensively
studied in the literature [
In previous work, a comparative evaluation of available ATE tools has been
performed [
Many of the tools analysed are web based applications: Translated.net4, TermoStat5 and FiveFilters6, for instance, are free tools that can be accessed online.
However, some of the issues found in the evaluation above mentioned include the extraction of stop words, visualisation of terms only in the website (not downloadable) and di culties to nd terms from speci c domains.
Other tools o er payable services, such as SketchEngine7, which is a
sophisticated tool with a good performance and additional services [
On the other hand there are other downloadable tools available, such as TBXTools8 and TermSuite9. The rst one has been developed to o er domain and language independent services, thus, implemented patterns might be too general; and the latter does not extract compound terms, the main feature of legal jargon.
This contribution uses the open source library JATE10, since it integrates the most important terminology extraction algorithms and can be integrated in a local Solr indexer. Furthermore, it can also process large corpora and be extended for other languages or for di erent purposes, such as the Spanish legal domain. 4
In the rst place, subsection 4.1 describes the extension of JATE for Spanish language and the initial set of patterns applied. Secondly, subsection 4.2 contains an analysis of the features of legal terminology, speci cally within Spanish labour law corpora. Afterwards, subsection 4.3 proposes the Spanish language patterns to be con gured in JATE based on the analysis above mentioned. Lastly, subsection 4.4 contains a description of the extraction tests performed. 4.1
JATE has implemented ten algorithms: TTF, ATTF, TTF-IDF, RIDF, CValue, X2, RAKE, Weirdness, GlossEx and TermEx.
In these experiments, only Cvalue [
JATE relies on the OpenNLP library11 for tasks such as tokenisation,
sentence segmentation, part-of-speech (POS) tagging and chunking, being the last
two the most signi cant to process Spanish documents. For the POS tagging in
Spanish, a model trained for version 1.3 has been adapted to the latest version.
The di erent POS tags are coded with Cast3LB format [
In order to create the Spanish patterns, previous work with English tagsets
has been taken as a reference [
The initial tag patterns used for the rst extraction tests against a general corpus were structured as in Figure 2.
Based on this gure, examples of the rst patterns presented are: { (nbNCnb): Noun Common (e.g. regulation) { (nbNCnb) (nbNCnb) : Noun Common + Noun Common (e.g. family background ). { (nbAQnb) (nbNCnb) : Adjective Qualifying + Noun Common (e.g. national jurisdiction).
Some peculiarities of the patterns showed in Figure 2 are, for instance, that
only common nouns have been considered to be extracted as simple terms. Verbs
and adjectives of general knowledge tend to be less relevant when building a
domain independent vocabulary. Hence, their tags have not been added to the
patterns to avoid noise generation.
The patterns exposed in the previous section were intended for general
information extraction. However, legal language has its own peculiarities that need to
be considered [
Legal terms of art: technical words with exact meaning that cannot be replaced by other terms (e.g. comodato, meaning \bailment") Legal jargon: terms and expressions used by lawyers, often archaic and obsolete words (e.g. lo antedicho, meaning \the aforesaid") Terms from the general language with a di erent meaning in the legal domain (e.g. furnish, meaning \to provide something or send something") These terms should be identi ed since, although in a general context they may have little signi cance, they are relevant in a legal context.
The labour law corpus used for this experiment is composed by 20 documents containing information on national company agreements from di erent regions in Spain. Keeping the previous characteristics in mind, this corpus was analysed before the extraction and the following considerations regarding legal terminology were raised: { The major part of the terms are multiword expressions (e.g. convenio colectivo, Bolet n O cial del Estado, grupo profesional ) { Many of these multiword expressions are built with prepositions and contractions connecting their elements (e.g. comite de empresa, prevencion de riesgos laborales, estatuto del trabajador ) { Some types of words that in a general context are not considered terms, such verbs and adjectives, in the legal domain do need to be extracted as terms since they have speci c meaning (e.g. vigente, enunciativo, devengar, retribuir ) { For the purposes of these experiments, proper names are not considered legal terms and must be avoided in the extraction stage (e.g. Comunidad de Madrid, Principado de Asturias, empresa Hermanos Fernandez ) { Similarly, URLs, numbers and dates must also be kept out from the resulting termlists (e.g.\https://www.boe.es/") { Finally, terms including ordinal adjectives that in other jurisdictions may have a unique meaning, such as the \Third Amendment", do not exist in Spanish legislation, so these types of words should not be extracted 4.3
Thus, the rst version of the Spanish general patterns was modi ed accordingly. Some of the patterns were removed and other new tag con gurations were added:
Following the legal language analysis described in Section 4.3 and the morphological tagset showed in Section 4.1, some examples of the patterns added are: { (nbAQnb): Adjective Qualifying (e.g. vigente, meaning \in force") { (nbVMnb): Verb Main (e.g. devengar, meaning \accrure") { (nbNCnb) (nbAQnb) (nbSPnb) (nbDAnb) (nbNCnb): Noun Common + Adjective Qualifying + Simple Preposition + Determiner Article + Noun Common (e.g. Bolet n O cial del Estado, meaning \O cial Bulletin of the State")
On the other hand, other patterns have been removed to avoid noise generation, since they were not considered relevant for legal terminology: { (nbAOnb) (nbNCnb): Adjective Ordinal + Noun Common (e.g. Third Amendment ) { (nbNCnb) (nbNCnb): Noun Common + Noun Common. This pattern was deleted since it was observed that the POS tagger sometimes tags proper nouns as NC (common nouns), and the tool extracts structures that are not real terms, such as empresa Apple. { Patterns that link several common nouns or several adjectives have also been removed, since they are very extended in English but not in Spanish grammar which, normally, uses prepositions to separate each component of the term. 4.4
The testing corpus provided by Lynx partners contains 20 les comprised of 21,475 tokens in TXT format that have been automatically converted apart from PDF les from the labour law domain in Spanish. Two tests have been performed over the Lynx corpus, one per pattern set: { Extraction tests with Cvalue and TTF-IDF algorithms applying the General Spanish patterns.
Samples of the most relevant extracted terms are convenio colectivo (collective agreement), direccion de la empresa (company management), empresa (company) and trabajador (worker). { Extraction tests with Cvalue and TTF-IDF algorithms applying the Legal Spanish patterns.
Most of the main terms recognised with the general Spanish patterns still remain in the same positions with similar scores. However, legal patterns have introduced new relevant terms in the legal domain such as vacaciones por antiguedad (seniority holidays), asambleas convocadas por el comite (committeeorganised assemblies), documentos relativos a la liquidacion (liquidation documents), jubilado (retired) or disciplinarios (discplinary).
From the resulting lists of terms, sorted by relevance, the rst 200 terms12 have been considered as the most meaningful, since they have signi cance in their scores. In this context, \relevant terms" are those legal expressions that can be used to annotate and classify documents by topic or typology, this is, terms that represent the legal domain. Table 1 collects some of the new extracted by JATE, applying the Spanish Legal Patterns13. 12 http://doi.org/10.5281/zenodo.2385437 13 https://github.com/oeg-upm/terminology-extractor New terms
Cvalue trabajador tendra (worker will have) conocimientos adquiridos en el desempen~o (acquire knowledge during the performance) ley de prevencion de riesgos (risk prevention law) representacion legal de los trabajadores (legal representation of workers)
rma del presente (signing the present) texto refundido de la ley (combined text of law) bocmbolet n o cial de la comunidad (bocmo cial bulletin of the community)
entrada en vigor del presente (implementation of the present contract)
miembros del comite de empresa (members of the company committee)
comision mixta de interpretacion (mixed interpretation commission)
bolet n o cial de la comunidad (o cial bulletin of the community) bolet n o cial de la junta (o cial bulletin of the council)
Table 1. New terms extracted by the legal patterns New terms
TTF-IDF anonima [AQ] (anonymous)
exible [AQ] ( exible)
dura [AQ] (severe)
sanitario [AQ] (sanitary)
discontinuo [AQ] (discontinuous)
grave [AQ] (serious)
mixta [AQ] (mixed)
From this table, the following considerations can be educed: { New terms retrieved by CValue algorithm are comprised of multiword terms.
The algorithm has been enriched with more complex nominal chunks that are relevant in the corpus. However, for the TTF-IDF algorithm, the main results retrieved are based on the individual terms, in this case adjectives. { The tool presents some tagging mistakes: the term trabajador tendra has not correctly been extracted, since the last component is a verb and there is not such pattern in the set. From the con guration of the patterns, the tool has tagged this verb as a common noun or a qualifying adjective (see Figure 3). { Also, there might be some tokenisation mistakes in the source corpus, since the term bocmbolet n o cial de la comunidad has not been correctly extracted either. Another possibility is that the PDF to TXT conversion inserted mistakes such this one in the source les. The quality of the source le must be reviewed and improved. 5
Results present little variations: 6% of new terms applying legal patterns with Cvalue and 3.5% with TTF-IDF. This situation is mainly given by POS tagging mistakes: the tool tends to tag any unknown word as common noun since it is the most frequent type of word in texts. Thus, some proper names, URLs, verbs and adjectives are extracted as common nouns, spoiling the patterns. Training the model again with larger corpora from the domain would avoid many of these mistaken tags.
Also, not all the extracted terms are relevant for the legal domain. Some terms (e.g. sanitario, dura, mixta) do not belong to the legal terminology, so it would be required to generate a list with terms from the general usage to escape such extractions in future experiments.
On the other hand, many of the wrong extractions are caused by tokenisation
mistakes in the original corpus: clean and well structured source documents
would avoid this issue. Moreover, other libraries such as IXA Pipes [
Another immediate convenient step to improve and generate a sound set of legal patterns is to consult grammatical peculiarities of legal language with actual legal experts: lawyers, prosecutors, judges, law students, etc., since these professionals are the best knowledge source of this domain.
On the whole, these experiments have been performed to highlight the importance of developing terminology extractors without the need of using online platforms for domains that deal with sensible data that cannot be distributed to third parties.
Regarding the use of JATE, the best approach here is to use several algorithms (in this case, Cvalue and TTF-IDF) with di erent features and performances to get more comprehensive results and to get a better overview of terms of the corpus. For instance, in the experiments, the results have shown the importance of the nominal chunks in the corpus although there are adjectives and nouns that are importance by themselves.
Also, since the tool allows the implementation and creation of new
algorithms, another interesting experiment is the testing of other existing algorithms
such KEA [
Finally, although in these experiments only public data have been handled, a future line of work would be focused on identifying sensitive data to identify and remove named entities of persons and organizations in a given corpus to treat personal and private data. Furthermore, entity linking processes of the retrieved terms are also considered in order to validate accuracy of the terms, and search for related terms in relevant knowledge bases. 14 http://www.lsi.upc.es/ nlp/tools/parole-sp.html