TIA PFIA 2019 Terminologie et Intelligence Artificielle (atelier TALN-RECITAL & IC) - IRIT
←
→
Transcription du contenu de la page
Si votre navigateur ne rend pas la page correctement, lisez s'il vous plaît le contenu de la page ci-dessous
TIA Terminologie et Intelligence Artificielle (atelier TALN-RECITAL & IC) PFIA 2019
Table des matières Thierry Hamon et Natalia Grabar. Éditorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 . Comité de programme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Claudia Lanza et Béatrice Daille. Terminology systematization for Cybersecurity domain in Italian Language . . . . . . . . . . . . . . . . . . . . . . 7 Tsanta Randriatsitohaina et Thierry Hamon. Identification des catégories de relations aliment-médicament . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Hamid Mirisaee, Eric Gaussier, Cedric Lagnier et Agnes Guerraz. Terminology-based Text Embedding for Computing Document Similarities on Technical Content 31 Nadia Bebeshina-Clairet et Sylvie Desprès. Exploiter un réseau lexico-sémantique pour la construction d’ontologie . . . . . . . . . . . . . . . . . . . . . . . . . 43 Kyo Kageura et Long-Huei Chen. Entropic characterisation of termino-conceptual structure : A preliminary study . . . . . . . . . . . . . . . 55 Frédérique Brin-Henry, Rute Costa et Sylvie Desprès. TemPO : towards a conceptualisation of pathology in speech and language therapy . . . . . . . . . . . 69
Éditorial
Présentation de l’atelier
Le groupe TIA (Terminologie et intelligence artificielle) a fondé la conférence TIA qui se tient depuis 1995 :
TIA’95 (Villetaneuse), TIA’97 (Toulouse), TIA’99 (Nantes), TIA’01 (Nancy), TIA’03 (Strasbourg), TIA’05
(Rouen), TIA’07 (Nice), TIA’09 (Toulouse), TIA’11 (Paris), TIA’13 (Villetaneuse), TIA’15 (Grenade). Fondé
sous l’égide de l’AFIA en 1994, devenu groupe AFIA/GdR-I3, TIA a rassemblé des chercheurs en linguistique,
en traitement automatique des langues et en ingénierie des connaissances.
En 2019, le groupe TIA a fêté ses 25 ans.
Nous avons profité de la plate-forme AFIA, qui a permis de joindre en 2019 les deux conférences, TALN-
RECITAL et IC, représentant les deux principales communautés scientifiques des membres du groupe TIA, pour
organiser un atelier commun.
Dans la continuité de l’esprit de la conférence TIA, cet atelier interdisciplinaire a eu pour vocation d’inviter
les chercheurs en terminologie, linguistique et IA (en particulier, en traitement automatique des langues, en
ingénierie des connaissances et en web sémantique) afin de s’interroger sur les liens renouvelés entre ces disci-
plines. Cela a semblé en effet important à un moment où les approches à base de réseaux de neurones et de
modèles statistiques des langues connaissent un succès retentissant.
Thèmes de l’atelier
Cette édition 2019 s’est focalisée sur les connaissances nécessaires à l’étude des terminologies et des langues
en complément aux traitements automatiques et au deep learning.
TIA 2019 a sollicité des contributions significatives mais aussi des travaux en cours ou plus ciblés notamment
autour des thèmes suivants :
— Acquisition et enrichissement de terminologies et d’ontologies
— dans un contexte monolingue ou multilingue
— pour des langues peu-dotées
— Terminologie selon l’expertise des utilisateurs
— Terminologie et ontologie
— Fusion de terminologies et d’ontologies
— Peuplement d’ontologies
— Interopérabilité sémantique
— Évolution de terminologies et d’ontologies
— Utilisation de terminologies et d’ontologies
— pour l’annotation sémantique
— pour le TAL
— pour l’éducation
— dans des applications industrielles
Les actes rassemblent six contributions proposées par des chercheurs français, italiens, portugais et japonais.
Ces publications portent sur des données textuelles en anglais, français et italien, et couvrent un large éventail de
travaux allant de l’acquisition de terminologies à la construction d’ontologies. De même, les approches proposées
restent assez classiques et s’appuient sur des ressources existantes (FrameNet, réseaux lexico-sémantiques), des
représentations sous forme de graphes, des mesures statisques et l’apprentissage automatique. Nous espérons
que l’atelier a offert suffisamment de possibilités d’avoir des discussions intéressantes et a offert un contexte
fructueux pour l’émergence de nouvelles problématiques ou de nouvelles collaborations.
Les organisateurs remercient le comité de programme pour leur important travail de relecture.
Thierry Hamon et Natalia Grabar
TIA@PFIA 2019 4Comité de programme
Organisateurs de TIA
— Natalia Grabar (CNRS UMR 8163 STL, France)
— Thierry Hamon (LIMSI-CNRS & Université Paris 13, France)
Comité de programme de TIA
— Nathalie Aussenac (IRIT, Universite Paul Sabatier, France)
— Jean Charlet (AP-HP, LIMICS/INSERM, Sorbonne Université, Univ. Paris 13)
— Anne Condamines (CLLE, CNRS et Université Jean Jaurès, France)
— Patrick Drouin (OLST, Université de Montréal)
— Pamela Faber (Universidad de Granada, Espagne)
— Kyo Kageura (Interdisciplinary Initiative in Information Studies/Graduate School of Education, The
University of Tokyo, Japon)
— Marie-Claude L’Homme (OLST, Université de Montréal, Canada)
— Adeline Nazarenko (LIPN, Université Paris 13 – Sorbonne Paris Cité, France)
— Aurélie Picton (TIM/FTI, Université de Genève)
— Thierry Poibeau (Lattice-CNRS, France)
— Pascale Sébillot (IRISA & INRIA / INSA Rennes, France)
— Sylvie Szulman (LIPN, Université Paris 13, France)
— Yannick Toussaint (LORIA, Université de Lorraine, France)
— Mathieu Valette (ERTIM, Inalco, France)
— Pierre Zweigenbaum (LIMSI, CNRS, Université Paris-Saclay, France)
5 TIA@PFIA 2019Claudia Lanza et Béatrice Daille
Terminology systematization for Cybersecurity domain in Italian
Language
Claudia Lanza1, 2 Béatrice Daille3
(1) University of Calabria (UNICAL), DIMES, via Pietro Bucci 87036 Arcavacata di Rende, Cs, ITALY
(2) PhD visiting student at LS2N Laboratory, Université de Nantes, FRANCE
(3) Université de Nantes - LINA CNRS UMR 6241, 2, Rue de la Houssiniere, BP 92208, F44322 Nantes Cedex
3, FRANCE
c.lanza@dimes.unical.it, beatrice.daille@univ-nantes.fr
A BSTRACT
This paper aims at presenting the first steps to improve the quality of the first draft of an Italian
thesaurus for Cybersecurity terminology that has been realized for a specific project activity in
collaboration with CybersecurityLab at Informatics and Telematics Institute (IIT) of the National
Council of Research (CNR) in Italy. In particular, the paper will focus, first, on the terminological
knowledge base built to retrieve the most representative candidate terms of Cybersecurity domain
in Italian language, giving examples of the main gold standard repositories that have been used to
build this semantic tool. Attention will be then given to the methodology and software employed
to configure a system of NLP rules to get the desired semantic results and to proceed with the
enhancement of the candidate terms selection which are meant to be inserted in the controlled
vocabulary.
K EYWORDS : alignment, cybersecurity, textual variants, thesaurus, term extraction
1 Introduction
In order to develop a means for terminological knowledge base, TBK (Condamines, 2018), the
construction of the corpus represents the first phase, this is followed by the extraction of candidate
terms and then by the construction of a relational system (Barrière, 2006). For what concerns the
systems on how to extract terminology, the webpage of the European Parliaments gives a detailed
overview of the available resources 1 . Among other studies that have given a state-of-the-art on the
different ways to extract information there is the analysis given by Bernier et al. (2012) for the gold
standard terms construction within a corpus-based methodology, or by Nazarenko et al. (2009) for the
evaluation of different terms extractors tools, or still other studies that proposed different methodology
to execute terminological extraction, like Aubin and Hamon (2006) or Loginova (2012) . Two of
the features to take into account in constructing a terminological knowledge base are accuracy and
precision with respect to the contextual requirements and target genre (Condamines, 2007). Another
key element for knowledge databases reliability is their representativeness with respect to a specific
field of knowledge (Condamines, 2018). According to Leech (1992), a semantic resource should
be considered representative if it contains an adequate level of information related to the domain of
reference.
Though a definite measurement or shared standardised calculation of the quantitative level of terms
1. http://termcoord.eu/discover_trashed/free-term-extractors/
term-extraction-tools/
7 TIA@PFIA 2019Terminology systematization for Cybersecurity domain in Italian Language a thesaurus should have (Caruso et al., 2014) doesn’t officially exist, the idea is to have as much information as possible about the domain of study. A useful method to improve the lexical richness could be that of Pastor and Seghiri (2010) according to which "new words will always be included in the corpus, a point is reached when the addition of more documents will not in practice bring anything new to the collection. In other words, it will not make it more representative because all the different situations, as well as the normal range of terminology in this particular field, have already been covered. ". Still, not always the larger size is a signal of a good baseline for a reliable corpus (Leech, 1991). 2 Resources for the Cyber domain The main purpose of the research activity related to the construction of a thesaurus for Cybersecurity domain in Italian language was that of providing a web resource able to structure the technical information proper to this area of study from a semantic point of view. In order to achieve this goal, an accurate study of the existent repositories in Cybersecurity domain has been carried out. Cybersecurity is a multidisciplinary field of knowledge. Its terminology is highly linked to the world of Information and Communication Technology (ICT) and its sub-areas 2 , but it’s also related to the legislative systems and regulations of the reference countries. For its being a combination of different fields of study, the corpus creation followed this orientation and it has been made up of documents both deriving from legal world and from sector-oriented magazines. To build up a source corpus that could include this convergence the gold standards have been gathered in order to start with a systematization of specific information. Indeed, these sources, the gold standards (Terryn et al., 2018), give a reliable content that has to be examined as to offer a comprehensive terminological database that has to be processed to create an accurate knowledge representation. The main difficulty in designing the source corpus at first was that English sources in this technical domain seemed more numerous than the Italian ones, therefore a process of alignment between these two languages appeared necessary (Daille, 2012). The main gold standards that have been exploited as the starting points to evaluate the results of terms extraction are in English : NIST 7298 (Kisserl, 2013) and ISO 27000 (2016) , and they have been manually translated to Italian. The goal was to have a Italian semantic resource, as well as to enhance the list of terms included in these latter standards related to Cybersecurity since the collaboration with experts made evident that there was a a level of abstractness in the terminology contained in them. Stating that a level of abstractness was present in the lists of terms the standards provide, means that sometimes words like "Activity", "Analysis" or "Objective" didn’t, in accordance with experts of Cybersecurity domain, give an high informative additional value to the terminological structure, then not needed for guiding the domain terminology comprehension that is characterized by a marked level of technicisms. The purpose was, indeed, to provide a precise semantic source that could organize in a piloted way the terminology of this technical domain. Among the Italian resources used to populate the source corpus there are the glossary edited by the Presidency of Ministers "Glossario Intelligence" 3 and the guideline book about Cybersecurity (Baldoni et al., 2018) which has been launched in 2015 during ITASEC 4 conference. If the latter provides a very well-comprehensive book that explains in details the most relevant elements related to Cybersecurity domain giving also examples of the key episodes occurred in years, the first is conceived as a tool that gives a list of the most important administrative and politics-oriented words 2. http://uis.unesco.org/en/glossary-term/ict-related-fields 3. https://www.sicurezzanazionale.gov.it/sisr.nsf/quaderni-di-intelligence/ glossario-intelligence.html 4. https://www.itasec.it/2018.html TIA@PFIA 2019 8
Claudia Lanza et Béatrice Daille
linked to Cybersecurity field of knowledge. Looking at the English language, instead, there is a large
number of databases that deal with Cybersecurity world. Starting from the standards ISO 27000 :2016
and NIST 7298 (Kisserl, 2013), and moving to the list of vulnerabilities and cyberattacks provided by
MITRE 5 6 . It’s worth mentioning also NICCS glossary 7 and the vocabulary offered by ENISA 8 . All
of them have represented sources that have guided the process of creating a thesaurus, because it’s
only by knowing which can be the specialized contexts in which the tool will be used that a semantic
resource like a thesaurus can check if the terminology obtained by the lexical extraction is accurate
(Condamines, 2018).
The main objective of this activity is to realize an image of the domain realizing a semantic tool that
could include the knowledge of the field. In order to achieve this first objective, the collaboration
with experts proved to be fundamental. It’s been thanks to their supervision that the first draft of the
thesaurus could have been structured in main categories. Currently this semantic resource is present
among the services on the website developed by the CybersecurityLab in Pisa (Italy) 9 ; this online
platform was created to orientate the comprehension the Cybersecurity domain, and the thesaurus
contains 245 terms, almost all of them with their definition extracted by the corpus documents.
This paper is focused on the way to improve the term extraction accuracy and to extend the first version
of the Italian thesaurus for Cybersecurity by having different rules beyond the lexical engineering
process.
3 Cyber corpus
Following the words of Bowker (2002), "A corpus is not only a random collection of texts [...]. Rather,
the texts in a corpus are selected according to explicit criteria in order to be used as representative
sample of a particular language or subset of that language".
In order to have a set of documents that should characterize the source corpus, choosing specific
parameters for selecting them it’s a very important phase. Linguistics of corpora (Pearson, 1998)
suggests to take in consideration documents that are related to present age, better if they share the
same time range and are all written in the same language, including produced in a same national area.
The domain of Cybersecurity is a field that undergoes towards continuous changes over time (Conda-
mines, 2018), and its lexicology gets involved too. Having a semantic source that can monitor the
terminology of a technical domain can be helpful both for common users and for experts as to follow
the lexical internal system and be obliquely updated with terms modifications in time.
For the realization of the Italian thesaurus for Cybersecurity, the hierarchy of sources in law has been
followed (Zagrebelsky, 1984). According to this upper level classification of reference sources from
which a construction of corpora should start, the first documents to take in consideration related to the
legal world, are, firstly, national, then regional and finally local. Subsequently, the corpus has included
divulgative documents both in digital and paper-native forms (i.e., sector oriented magazines).
Therefore, for the development of the source corpus for the Italian thesaurus of Cybersecurity, legisla-
tive datasets have firstly been taken into account. The main portals on which this kind of information
has been retrieved were the European official websites where it’s possible to freely download the laws
of interest, i.e., EUR-lex, Altalex. Each of these online platforms allows the selection of many filters
related to the classification of legislative sources (Penal, Civil, Administrative Code), the time range,
the keywords used to search inside the database. Among these documents that have made up the
5. https://capec.mitre.org/
6. https://cve.mitre.org
7. https://niccs.us-cert.gov/about-niccs/
8. https://www.enisa.europa.eu/topics/threat-risk-management/risk-management/
current-risk/risk-management-inventory/glossary
9. https://www.cybersecurityosservatorio.it/it/Services/thesaurus.jsp
9 TIA@PFIA 2019Terminology systematization for Cybersecurity domain in Italian Language
source corpus there are also technical documents, such as reports or guidelines published by official
institutions’ portals that are monitoring the Cybersecurity sphere in the Italian field, i.e. CERT 10 .
Papers or scientific articles have been excluded at the moment because the terminological texture
might have been too discursive and the tone of the sentences too adverbial. Cybersecurity is a domain
that owns specific technicism, in this way the first selection of documents has been structured, and
gradually a list of the main legislative Italian documents related to Cybersecurity over the latest years
has been set up. The selection of the texts to be inserted in the collection of document that would have
made up the source corpus has been mainly based on the retrieval of the most important laws, norms,
legislations, regulations about the domain of Cybersecurity. The importance of certain documents
instead of others has been given by some filters related to :
— time range ; documents taken into account were ranged around the latest years, minimum the
latest seven ;
— language : only Italian documents have been considered for the analysis ;
— contexts : national, european and regional laws have been analysed.
Nevertheless, given that the scope of the thesaurus for Cybersercurity was to include as much
information as possible on this domain, legislative documents, which constituted a relevant part of
the corpus, seemed not enough powerful to cover the knowledge specificity domain’s spectrum. It is
for this latter reason that also sector-oriented magazines have been employed for the data population
in the corpus to be processed. Table 1 summarizes the size of the Italian Cybersecurity corpus by the
number of documents and the number of words that have been used for the building of the Italian
thesaurus of Cybersecurity. The overal size of the corpus is 563 documents with 11 806 558 terms
contained in them.
Number of documents Number of terms
Laws 249 8 918 209
Magazines 315 7 640 732
TABLE 1 – Features of the Italian corpus of Cybersecurity
4 Methodology
For the term extraction process the starting reference schematization has been that contained in
the thesaurus whose terms and semantic relations have been evaluated and validated during several
months of collaboration with the experts of this domain. The first draft of the Italian thesaurus for
Cybersecurity has as its peculiarity that of having been written in Italian official language, and this
trait could be considered as a innovative investigation on the development of a semantic resource for
organizing the information on Cybersecurity field of knowledge. Terms which are currently present in
the Italian thesaurus are partly represented by the most frequent terms given by T2K 11 term extraction
and partly structured according to the information progressively given by experts during the important
phase of direct collaboration. Up to now, the first comparative structure to check the granularity
obtained with TermSuite 12 is the first draft of the Italian thesaurus for Cybersecurity which had
been already evaluated from the experts. The terminological process started at the beginning by
selecting the most frequent terms, according to TF/IDF formula, among the list of terms given by the
semi-automatic software extractor T2K. These terms have gone through a careful validation process
with the experts of the domain with whom a selection of the preferred terms has been settled down.
10. https://www.certnazionale.it/
11. http://www.italianlp.it/demo/t2k-text-to-knowledge/
12. http://termsuite.github.io/
TIA@PFIA 2019 10Claudia Lanza et Béatrice Daille
This paper aims at describing the further steps that anticipate the improvement of the current semantic
structure of the thesaurus. The ongoing goal is, hence, to enhance the current thesaurus structure,
made up of 245 terms, through the variations detected by TermSuite in the source corpus and with
Pke library (Boudin, 2016) for clustering the keyphrases. In the seek of aligning the information
included in the thesaurus for Cybersecurity, the passages hereafter described will present :
1. The mapping procedure executed between the terms of the thesaurus and the standards of
Cybersecurity, i.e. Nist and ISO 27000 :2016, showing how many of them appear in these
reference resources and why up to now several of them have not been taken into account ;
2. The terminological extraction by exploiting the semi-automatic tools :
— T2K will be the first to be described as the first that has been used to extract the first
terminological dataset ;
— TermSuite will be presented as a software that helped in enhancing the structuring of the
semantic relations by using variants ;
— Pke as the library that with its models, TopicRank and Multipartite Rank, supported the
definition of topical information about the domain. In detail, for what concerns Multipartite
Rank, the first executions have been tested on the main four macro-categories included in
the Italian thesaurus for Cybersecurity that have been validated by the group of experts :
Cybersecurity, Cybercriminality, Cyberbullism, Cyber Defence.
4.1 Mapping
The process of checking the candidate terms has been firstly executed by mapping the term records
obtained with T2K with the words contained in the list of NIST and ISO 27000 :2016 vocabularies
that can best suite the needs of the group of experts who played a relevant role for the construction
of the Italian thesaurus. The Italian thesaurus for Cybersecurity also underwent through a process
of validation in comparison with the main gold standards, i.e. NIST and ISO 27000 :2016, the
terminological data set inside it can be reasonably conceived as a reliable structure with which
juxtapose the results from the terms candidates extraction software. The lists of terms contained in
the standards NIST and ISO 27000 :2016 have been manually translated with the support of terms
databases in IATE system 13 in order to retrieve the information using the same language in input as
we worked for the construction of the thesaurus on Cybersecurity in Italian. Through the mapping
with these standards, the check was aimed to verify the presence or not of the selected terms given by
T2K extraction. The further passage dealt with coverage of the terms contained in the corpus used to
build the semantic resource with respect to the standards, as to demonstrate how the terminology given
by the texts gathered in the corpus could be overlapped with the list of terms in the standards.Table 2
summarises the scores given by the comparison. As stated in section 2, several general words like
"Analysis" have been put aside for the moment because too vague for the purposes of providing a
semantic resource to structure the technicisms of the domain.
Nist Match ISO match
Thesaurus 75/1299 17/88
terms
TABLE 2 – Matches between the terms contained in the thesaurus and the ones contained in the
Standards for Information Security.
13. https://iate.europa.eu/home
11 TIA@PFIA 2019Terminology systematization for Cybersecurity domain in Italian Language
4.2 Term Extraction software
Once defined the documents that should have made up the source corpus, the candidate term extrac-
tion begun. For this phase the decision was to use several tools that presented different features. For
instance, two of the principal software that have been exploited to retrieve the terminology are Text
To Knowledge (T2K) (Dell’Orletta et al., 2014) and TermSuite (Cram & Daille, 2016).
This activity refers to an ongoing process of selection of the more adequate tool from which termino-
logy on this domain has to be taken into consideration. For the first development of the thesaurus in
Italian language for Cybersecurity, T2K has been used, but, as it will be better specified in the nexts
paragraphs, some of its internal restrictions limited the visualization of different types of variants.
Nonetheless, for the purposes of the first draft of the Thesaurus, this term extractor software proved to
give reliable results that have been approved by the consensus of the expert on this specified field. The
selection of terms amongst the ones given in output by T2K has been based both on the first scores
sorted according to TF/IDF formula, and on the schemas owned by the experts to frame the main
Cybersecurity concepts, i.e. the head term Cyber accompanied by several key concepts of this field
of study, such as "hygiene", "espionage", "threat". In order to have a better systematization of the
Italian thesaurus for Cybersecurity to be finalized also in agreement with the community of experts,
another resource that is giving interesting results is Pke (Boudin, 2016). This library dedicated to
keyphrases extraction gave as output a good grouping structuring of the information contained in the
source corpus that can guide the orientation of certain macro-categories inserted or to be inserted in
the thesaurus with other concepts with which they are related.
While there are some fundamental differences between term extraction and keyword extraction
approaches (Lossio-Ventura et al., 2014), we wanted to test their faculty to propose other types of
terms selected according to their keyness properties at the document level.
4.2.1 T2K
T2K (Dell’Orletta et al., 2014) is native Italian extractor, has been created by the Italian Computational
Linguistics Group at CNR in Pisa. In addition to extract domain-specific terminology and phrases, it
organizes and structures the set of extracted terms and phrases into taxonomical chains, provides PoS
tagging for the corpus as well as a dependency parsing overview.
Its main advantages are pretty much related to the fact that is a software developed by Italian spoken
experts, so the rules beyond the system result to be specifically orientated to the source language of
the thesaurus for Cybersecurity. Another benefit is that it can apply a contrast function with another
reference vocabulary or a list already processed, as to make two terminilogical sets comparable for
statistical analysis.
For what concerns one of its cons, even though it’s a corpus oriented extraction tool, it doesn’t
present a very wide customization intervention in the grammar rules. However, from the beginning
it’s possible to decide which kind of lexical chains is the extraction meant to provide, as well as the
threshold frequency range and the length of terms.
The candidate terms extraction resulted to be quite different in using the two term extractor software.
The results are presented in Table 3. In detail, given the limited range of modifications T2K allows
to do, the total numbers of entries (single terms and MTW) was 446 325, a very high number that
reveals how many noise can be present among the records. That is probably linked to the fact that
T2K doesn’t show the possibility to insert a personalised stopword list or a set of variation rules
like TermSuite. The outnumbered results in T2K have been obtained by applying the following
configuration structure :
1. the frequency has been set up to level 1 ;
2. the semantic chains have been defined as Common Nouns (S) followed by an Adjective (A),
TIA@PFIA 2019 12Claudia Lanza et Béatrice Daille
T2K TermSuite
number of candidates 446 325 16 641
TABLE 3 – Terminology extraction : number of candidate terms extracted by each tool.
or/with a a Preposition (E), Articulated Preposition (EA), Common Noun (S), and ending
with a Common Noun (S) or an Adjective (A) ;
3. Maximum length terms has been fixed at 8.
In TermSuite, the overall results on the same corpus, made up of almost 563 documents, resulted to be
16 641, with a discrepancy of 429 684 units. Even though the results have been lower than T2K, the
application of Italian rules beyond the system, the customization of stopword lists, the configuration
of prefix banks and derivational datasets, allowed to obtain more sophisticated results.
4.2.2 TermSuite
TermSuite (Cram & Daille, 2016) is not an Italian native extractor but, as it is multilingually designed,
it could be extended or adapted to another language. TermSuite computes the termhood and the
unithood of a term candidate as defined by Kageura and Umino (1996). It adopts the two core steps
of the terminology extraction process :
1. Identification and collection of term-like units in the texts, mostly a subset of nominal phrases ;
2. Filtering of the extracted term-like units that may not be terms, syntactically or terminologi-
cally ; sorting of the candidate terms according to their unithood, their terminological degree
and their most usefulness for the target application.
Among its main features, the variants recognition, that could help for relation detection, proved to be
very helpful in terms of enhancing the first structure of the thesaurus, because acting on the linguistics
rules the output can be more accurate and customized. Other features beside the patterns recognition,
are for example the definition of different typologies of frequency (i.e., Document Frequency, TF-IDF,
General Frequency Norm).
4.2.3 TermSuite customization for Italian
One of the main advantages of TermSuite has been that of creating customized patterns that enabled
the extraction of desired lexical outputs.
The morphology package was the one that has been mostly edited, starting from the prefixes and
suffixes, acting on the derivation bank in order to adjust the rules to the Italian ones. The number
of grammar rules in Italian does not differ from the French ones. For the purposes of the ongoing
enhancing systematization for the Cybersecuirity Italian thesaurus, the variation rules made in French
language appeared strictly similar to what might have been the Italian ones. A core grammar of
variants is instantiated for five languages including French and Spanish languages which gather
morphological and syntagmatic rules and three main categories of variants : denominative, conceptual
and linguistic variants (Daille, 2017). Table 4 summarises the number of rules by language of the
default TermSuite Grammar.
To enhance the terminological extraction for the Italian thesaurus for Cybersecurity, a structuring of
Italian variation rules has been arranged. Therefore, in the extraction given by TermSuite, to obtain
different form of variations, syntactic as well as morphological, an alignment with the Italian grammar
rules proved to be necessary. Restructuring the patterns for Italian meant the specification of the most
common semantic chains the Italian language owns, the base-terms (Daille, 2003). Indeed, in Italian
language the main basic syntactic structures are made up of : Noun Adjective sicurezza informatica
13 TIA@PFIA 2019Terminology systematization for Cybersecurity domain in Italian Language
Spanish French
Denominative variants Morphological 2 2
Syntagmatic 4 5
Conceptual variants Morphological 2 4
Syntagmatic 17 13
Linguistic variants Morphological 0 2
Syntagmatic 15 11
Total Morphological 4 8
Syntagmatic 36 29
Number of rules 40 37
TABLE 4 – TermSuite grammar : number of rules for Spanish and French, by structure type, and by
morphological or syntagmatic nature.
(cybersecurity), Noun(Prep(Det))Noun titolare del trattamento (treatment manager)
Each of these basic patterns of base-terms can be modified by using the head part of these latter and
clustering the occurrences found in the source texts.
To give some examples of how the variations related to Cybersecurity domain in Italian language
appear after having activated the variation rules in TermSuite, we observed the three types of variants :
1. Denominative variants
— Composition : npn : hacker del telefono (phone hacker) → na : hacker telefonico
— Lexical reduction : npna : indirizzo di posta elettronico (mail address) → na : indirizzo
elettronico (electronic address)
2. Conceptual variants
— Derivation : na : contenuto legale (legal content) → na : contenuto illegale (illegal content)
— Expansion : na : fascicolo sanitario (personal health record) → naa : fascicolo sanitario
elettronico (electronic personal health record)
3. Linguistic variants
— Graphical segmentation : cyber-security ↔ cybersecurity
— Coordination : npn : sicurezza della rete (network security) → npncpn : sicurezza della
rete e del sistema (network and system security)
Denominative variants in the thesaurus shall be considered as units to be incorporated as synonymous
variants of terms which have been approved by the experts of the domain. Having a list of terms
accompanied by the synonymy form is a very helpful base from which begin to structure the network
system of the semantic relationships in the thesaurus, as well as knowing the conceptual variants that,
by expanding the terms output through the insertion of words with which they appear in the source
corpus, can provide a better frame of the information inside the documents of the corpus as to decide
which can be the best to be inserted in the thesaurus. For what regards the linguistic variants, such as
coordination or the graphical segmentation, all represent a valid decision-making system to recognize
which can be, among the entries given by a candidate term extraction, the preferred ones approved by
the experts communities.
4.2.4 Pke
Another means that has been tested is the library Pke (Boudin, 2016). Pke library is purely statistical,
so it’s by nature multilingual and is used for keywords extraction. It is not purely term extraction,
it’s a document oriented extraction. It was used to extract information about the topic terms, single
TIA@PFIA 2019 14Claudia Lanza et Béatrice Daille
or multi-word, that characterised the documents contained in the corpus as to use them in order to
check the thesaurus structure of semantic relationships. Pke algorithms have two steps : a keyphrase
candidate extraction step followed by a refinement step. Sequences of adjacent words, restricted to
nouns and adjectives, are considered as keyphrase candidates. Refinement step orders the keyphrase
candidates according to various statistical or graph-based methods.Indeed, this means to retrieve
keywords from a source corpus proved to be very useful in detecting the main semantic areas.The
following are example taken from the output given by executing TopicRank algorithm, where the
first are the main topics retrieved in the corpus alongwith their weight size :’sistema informatico’
(informative system), ’reato’(crime),’dati(data),’accesso illegittimo’(illegitimate access), ’accesso
abusivo’(abusive access),’rete’(textitnetwork), ’sito’(website),’server’(server).
We then tried to perform the Multiplartite Rank (Boudin, 2018) and the results proved to be better for
the purposes of helping in structuring the categories present in the current thesaurus. The following
are examples of the several clusters sorted by relevance given by this library that could depict how
the terms can be correlated using the keyphrases extraction logic. These representative cases are the
results given by searching the way the main four macro-categories of the current structure of the
thesaurus, i.e. Cybesecurity, Cybercriminality, Cyberbullism, Cyber Defence, are organized by this
model :
1. Cybersecurity :
— From a technical document : sicurezza (security), sistema (system), attacchi cibernetici
(cyber attacks), dati personali (personal data), cybersecurity (cybersecurity)
— From a divulgative document : here it’s intersting to note that the term is written in a
separated way, i.e. Cyber security, that could be a signal that the divulgative terminology
is different from the technical one :
cyber security (cyber security), sicurezza (security), reti (networks), informazioni (infor-
mation), macchine intelligenti (smart cars),
2. Cybercriminality :
— From a legal document : cibercriminalità (cybercriminality), informazione (information),
lotta (fight), reti (networks), reato informatico (cybercrime)
— From a divulgative document : informazioni (information), sicurezza (security), criminalità
informatica (cyber criminality), comunicazione (communication), reti (networks)
3. Cyberbullism :
— From a divulgative document : "cyberbullism" gives just one occurrence prevenzione
(prevention), cyberbullismo (cyberbullism), sistema educativo (educational system), scuole
(school), formazione (education),
4. Cyber Defence :
— From a divulgative document : attacchi (attacks), attacchi cibernetici (cyber attacks),
minacce (threats), difesa cibernetica (cyber defence), reti networks)
— From a legal document : informazioni on-line (on-line information), sicurezza (security),
attacchi informatici (cyber attacks), difesa informatica (cyber defence), difesa (defence)
From these examples we can understand that Cyber defence is related to Cyber attacks which is in
turn connected with Threats, or that Cyber bullism is contextualized with the educational system and
schools. This schematization helped in structuring the information.
5 Candidate terms observation
To give an example of how different perform the two software, the following are some of the terms
present in the current thesaurus, which, for instance, have already undergone through a validation
15 TIA@PFIA 2019Terminology systematization for Cybersecurity domain in Italian Language
process by the experts of Cybersecurity domain, with their source extraction in T2K compared with
Termsuite. Sometimes, as the first example, TermSuite proved to provide a more informative result,
other times, T2K gave more outputs for one entry even not grouping it in a variation cluster, or, on
the contrary, it happened also that TermSuite presented a result that was not present inside the T2K
extracted terminology , but was very important for the experts of the domain, as the third example
shows. Finally an example of a new term provided by the Pke clustering.
1. Thesaurus term : "Attacchi a forza bruta"(BruteForce Attack), which is the more specific
term of "Cyber attacks mechanisms" :
— TermSuite a visualization by variation : Term na : attacco bruto → Denominative variants :
npna : attacco a forza bruto and npna : attacco di forza bruto ;
— T2K : Prototypical form : attacco a forza bruta, Lemma of term :attacco a forza bruto ;
— Pke : this term is not specifically present, but if we look for Attack we have different score
according to if its a cluster of ten keyphrases in a legal document or in a sector-oriented
issue of the magazines :
(a) Legal document :Sicurezza (security),sistema (system), attacchi cibernetici (cyber at-
tacks), dati personali (personal data), cybersecurity (cybersecurity), attacchi (attacks),
rete (network), protezione (protection) ;
(b) Divulgative document : siti(website), http (http), pagina (page), attacco (attack), sis-
tema (system), posta elettronica (e-mail), etica hacker (hacker ethics), virus (virus) ;
2. Thesaurus term : "Phishing", which is a more specific term of "Spam", is given in two
different outputs :
— Termsuite : n : phishing and npn : messaggio di phishing (phishing message) ;
— T2K : the terms is under 176 terminological units, among these latter the following
are the first five representative results of the extraction. Prototypical forms : tecniche di
phishing (phishing techinques), mail di phishing (phishing mail), spear phishing, attacchi
di phishing (phishing attacks), siti di phishing (phishing websites) ;
— Pke : not present
3. Thesaurus term : "Cavalli di troia"(Trojan Horses), which is a more specific term of
Dipendent Software, that are, in turn, more specific terms of Malicious Software :
— T2k : not present
— TermSuite : TermSuite retrieved both "Trojan" alone and "Trojan-horses", which is more
precise and accurate for the purposes of the thesaurus on Cybersecurity ;
— Pke : not present
Pke new term : threat intelligence whose clusters are : threat intelligence (threat intelligence),
controspionaggio informatico (cyber counter-espionage), sistemi informativi (informative systems),
azioni (actions), intelligence (intelligence).
6 Conclusion
Enhancing the thesaurus terminological structure is a an ongoing perspective to be progressively
achieved. TermSuite extraction helped in obtaining more accurate results in terms of variants : having
a list of different ways to represent a terms results to be a valid supporting system to decide, alongside
the consensus of experts, the best entries with reference to accuracy and synonyms. Finally, Pke library
with Multipartite Rank gives several terminological groups which can help in detecting the semantic
interrelations among terms and work in future on the construction of thesaurus basic relationships by
employing its collection reasoning. as well as the inclusion of new terms given by the keyphrases.
TIA@PFIA 2019 16Claudia Lanza et Béatrice Daille
Références
AUBIN S. & H AMON T. (2006). Improving term extraction with terminological resources. In T.
S ALAKOSKI , F. G INTER , S. P YYSALO & T. PAHIKKALA, Eds., Advances in Natural Language
Processing (5th International Conference on NLP, FinTAL 2006), number 4139 in LNAI, p. 380–387 :
Springer.
BALDONI R., D E N ICOLA R. & P RINETTO P. (2018). Il Futuro della Cybersecurity in Italia :
Ambiti Progettuali Strategici Progetti e Azioni per difendere al meglio il Paese dagli attacchi
informatici. Laboratorio Nazionale di Cybersecurity (CINI) - Consorzio Interuniversitario Nazionale
per l’Informatica.
BARRIÈRE C. (2006). Semi-automatic corpus construction from informative texts. In L. B OWKES,
Ed., Text-Based Studies in honour of Ingrid Meyer, Lexicography, Terminology and Translation,
chapter 5. University of Ottawa Press.
B ERNIER -C OLBORNE G. . (2012). Defining a gold standard for the evaluation of term extractors.
In in Proceedings of the Eight International Conference on Language Resources and Evaluation
(LREC’12), p. 15–18.
B OUDIN F. (2016). pke : an open source python-based keyphrase extraction toolkit. In Procee-
dings of COLING 2016, the 26th International Conference on Computational Linguistics : System
Demonstrations, p. 69–73, Osaka, Japan : The COLING 2016 Organizing Committee.
B OUDIN F. (2018). Unsupervised keyphrase extraction with multipartite graphs. In Proceedings
of the 2018 Conference of the North American Chapter of the Association for Computational
Linguistics : Human Language Technologies, Volume 2 (Short Papers), p. 667–672, New Orleans,
Louisiana : Association for Computational Linguistics.
B OWKER L. & P EARSON J. (2002). Working with Specialized Language : A Practical Guide to
Using Corpora. London/New York : Routledge.
C ARUSO A., F OLINO A., PARISI F. & T RUNFIO R. (2014). A statistical method for minimum
corpus size determination. In 12es Journées internationales d’Analyse statistique des Données
Textuelles (JADT2014), Paris, France.
C ONDAMINES A. (2007). L’interprétation en sémantique de corpus : le cas de la construction de
terminologies. Revue française de linguistique appliquée, Vol. XII(2007/1), 39–52.
C ONDAMINES A. (2018). Terminological knowledge bases from texts to terms, from terms to texts.
In The Routledge Handbook of Lexicography. Routledge.
C RAM D. & DAILLE B. (2016). Terminology extraction with term variant detection. In Proceedings
of ACL-2016 System Demonstrations, p. 13–18, Berlin, Germany : Association for Computational
Linguistics.
DAILLE B. (2003). Conceptual structuring through term variations. In F. B OND , A. KORHONEN ,
D. M AC C ARTHY & A. V ILLACICENCIO, Eds., Proceedings ACL 2003 Workshop on Multiword
Expressions : Analysis, Acquisition and Treatment, p. 9–16 : ACL.
DAILLE B. (2012). Building bilingual terminologies from comparable corpora : The ttc termsuite.
DAILLE B. (2017). Term Variation in Specialised Corpora : Characterisation, automatic discovery
and applications, volume 19 of Terminology and Lexicography Research and Practice. John
Benjamins.
17 TIA@PFIA 2019Terminology systematization for Cybersecurity domain in Italian Language D ELL’O RLETTA F., V ENTURI G., C IMINO A. & M ONTEMAGNI S. (2014). T2K : a system for automatically extracting and organizing knowledge from texts. In N. C. C. C HAIR ), K. C HOUKRI , T. D ECLERCK , H. L OFTSSON , B. M AEGAARD , J. M ARIANI , A. M ORENO , J. O DIJK & S. P IPERIDIS, Eds., Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), Reykjavik, Iceland : European Language Resources Association (ELRA). ISO/IEC 27000 (2016). Information technology – Security techniques – Information security management systems – Overview and vocabulary. International Standard. K AGEURA K. & U MINO B. (1996). Methods of automatic term recognition : a review. Terminology, 3(2), 259–289. K ISSERL R. (2013). Glossary of Key Information Security Terms. National Institute of Standards and Technology. NISTIR 7298 Revision 2. L EECH G. (1991). The state of the art in corpus linguistics. London : Longman. L EECH G. (1992). Corpora and theories of linguistics performance. In J. S. ( ED .), Ed., Directions in Corpus Linguistics : Proceedings of Nobel Symposium, p. 105–122, Berlin : Mouton de Gruyter. L OGINOVA C LOUET E., G OJUN A., B LANCAFORT H., G UEGAN M., G ORNOSTAY T. & H EID U. (2012). Reference Lists for the Evaluation of Term Extraction Tools. In Terminology and Knowledge Engineering Conference (TKE), Madrid, Spain. L OSSIO -V ENTURA J. A., J ONQUET C., ROCHE M. & T EISSEIRE M. (2014). Towards a mixed approach to extract biomedical terms from text corpus. IJKDB, 4(1), 1–15. NAZARENKO A., Z ARGAYOUNA , H. ; H AMON O. & VAN P UYMBROUCK (2009). Evaluation des outils terminologiques : enjeux, difficultés et propositions. Traitement Automatique de la Langue (TAL), 50(1), 257–281. PASTOR G. C. & S EGHIRI M. (2010). Size matters : A quantitative approach to corpus representati- veness. Language, translation, reception. To honor Julio César Santoyo, p. 1–35. P EARSON J. (1998). Terms in Context. John Benjamins, Amsterdam. T ERRYN A. R., H OSTE V. & L EFEVER E. (2018). A Gold Standard for Multilingual Automatic Term Extraction from Comparable Corpora : Term Structure and Translation Equivalents. In N. C. C. CHAIR ), K. C HOUKRI , C. C IERI , T. D ECLERCK , S. G OGGI , K. H ASIDA , H. I SAHARA , B. M AEGAARD , J. M ARIANI , H. M AZO , A. M ORENO , J. O DIJK , S. P IPERIDIS & T. T OKUNAGA, Eds., Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan : European Language Resources Association (ELRA). Z AGREBELSKY G. (1984). Il sistema costituzionale delle fonti del diritto. Turin : UTET. TIA@PFIA 2019 18
Identification des catégories de relations aliment-médicament
Tsanta Randriatsitohaina1 Thierry Hamon1,2
(1) LIMSI, CNRS, Université Paris-Saclay, Campus universitaire d’Orsay, 91405 Orsay cedex, France
(2) Université Paris 13, 99 avenue Jean-Baptiste Clément, 93430 Villetaneuse, France
tsanta@limsi.fr, hamon@limsi.fr
R ÉSUMÉ
Les interactions aliment-médicament se produisent lorsque des aliments et des médicaments pris
ensembles provoquent un effet inattendu. Leur reconnaissance automatique dans les textes peut être
considérée comme une tâche d’extraction de relation à l’aide de méthodes de classification. Toutefois,
étant donné que ces interactions sont décrites de manière très fine, nous sommes confrontés au manque
de données et au manque d’exemples par type de relation. Pour résoudre ce problème, nous proposons
une approche efficace pour regrouper des relations partageant une représentation similaire en groupes
et réduire le manque d’exemples. Notre approche améliore les performances de la classification des
FDI. Enfin, nous contrastons une méthode de regroupement intuitive basée sur la définition des types
de relation et un apprentissage non supervisé basé sur les instances de chaque type de relation.
A BSTRACT
Identification of categories of food-drug relations.
Food-drug interactions occur when food and drug taken together cause unexpected effect. Their
automatic identification from texts can be viewed as a relation extraction task relying on classification
methods. Nevertheless, since these interactions are described in a very fine way, we are confronted
with a lack of data and a lack of examples per type of relation. To solve this problem, we propose
an efficient approach to cluster the relation sharing similar representation and to reduce the data
sparseness. Our approach increases the efficiency of the FDI classification. Finally, when grouping
the relations, we contrast an intuitive method based on the definition of the relation types, and an
unsupervised learning relying on the examples associated with the FDI relation type.
M OTS - CLÉS : Classification, Clustering, Extraction de relation, Textes biomédicaux.
K EYWORDS: Classification, Clustering, Relation Extraction, Biomedical texts.
1 Introduction
Bien qu’il existe des bases ou des terminologies recensant les connaissances d’un domaine de
spécialité, disposer d’informations à jour nécessite souvent d’avoir recours à l’analyse d’articles
scientifiques. Ce constat est d’autant plus vrai lorsque les connaissances à recenser ne sont pas déjà
présentes dans une base. Ainsi, si les interactions entre médicaments (Aagaard & Hansen, 2013) ou les
effets indésirables d’un médicament (Aronson & Ferner, 2005) sont répertoriés dans des bases telles
que DrugBank 1 ou Theriaque 2 , d’autres informations comme les interactions entre un médicament
1. https://www.drugbank.ca/
2. http://www.theriaque.orgIdentification des catégories de relations aliment-médicament et un aliment y sont très peu présentes. Elles sont souvent fragmentées et dispersées dans des sources hétérogènes, principalement sous forme textuelle. Afin de répondre à ces problématiques de mises à jour ou de recensement de ces informations, des méthodes de fouille de textes sont généralement mises en œuvre (Cohen & Hunter, 2008; Rzhetsky et al., 2009; Chowdhury et al., 2011). Dans cet article, nous nous intéressons à l’identification automatique de mentions d’interaction entre un médicament et un aliment (Food-Drug Interaction - FDI) dans des résumés d’articles scientifiques issus de la base Medline. A l’instar des interactions entre médicaments, une interaction entre un médicament et un aliment correspond à l’apparition d’un effet non attendu lors d’une prise combinée. Par exemple, le pamplemousse a un effet inhibiteur sur une enzyme impliquée dans le métabolisme de plusieurs médicaments (Hanley et al., 2011). Pour extraire ces informations des résumés, nous faisons face à plusieurs difficultés : (1) les mentions des médicaments et des aliments sont très variables dans les résumés. Il peut s’agir des dénominations communes internationales ou des substances actives de médicaments tandis que pour les aliments, il peut s’agit d’un nutriment, d’un composant particulier ou d’une famille d’aliment ; (2) les interactions sont décrites de manière assez fine dans le corpus annoté à notre disposition, ce qui conduit à un faible nombre d’exemples. ; (3) nous disposons de résumés annotés avec des interactions aliment/médicament mais ces annotations ne couvrent pas de manière homogène tous les types d’interaction et l’ensemble d’apprentissage est bien souvent déséquilibré. Nos contributions se concentrent sur l’extraction des FDI et l’amélioration des résultats des classifi- cations en proposant une représentation des relations qui prend en compte le manque de données, en appliquant une méthode de regroupement sur le type de relations, et en utilisant des étiquettes de groupe dans une étape de classification afin d’identifier le type de FDI. Après avoir présenté un état de l’art des méthodes d’acquisition de relations en corpus de spécialité (section 2), nous décrivons le corpus annoté que nous avons utilisé (section 3), notre approche pour regrouper les FDI annotées (section 4) et les expériences réalisées (5). Puis nous présentons et discutons les résultats obtenus (section 6) et nous concluons (section 7). 2 Etat de l’art Différentes approches ont été proposées pour extraire les relations à partir des textes biomédicaux. Certaines approches combinent des patrons lexico-syntaxiques avec un CRF pour la reconnaissance des symptômes dans les textes biomédicaux (Holat et al., 2016). D’autres approches génèrent automatiquement des données lexicales pour le traitement du texte libre dans les documents cliniques en s’appuyant sur un modèle alignement séquentiel multiple pour identifier des contextes similaires (Meng & Morioka, 2015). Toutefois, ces méthodes nécessitent des données bien fournies pour être efficace. Afin d’extraire les interactions médicament-médicament (DDI) (Kolchinsky et al., 2015) se concentrent sur l’identification des phrases pertinentes et des résumés pour l’extraction des preuves pharmacocinétiques. (Ben Abacha et al., 2015) proposent une approche basée sur du SVM combinant : (i) les caractéristiques décrivant les mots dans le contexte des relations à extraire, (ii) les noyaux composites utilisant des arbres de dépendance. L’union et l’intersection des résultats permettent d’obtenir des F1-mesures de 0,5 et 0,39 respectivement. Une méthode en deux étapes basée également sur le classifieur SVM est proposée par (Ben Abacha et al., 2015) pour détecter les DDI potentiels, puis classer les relations parmis les DDI déjà identifiées. Cette seconde approche permet d’obtenir des F1-mesures de 0,53 et 0,40 sur des résumés Medline et 0,83 et 0,68 sur des documents de DrugBank. (Kim et al., 2015) ont construit deux classifieurs pour l’extraction DDI : un classifieur binaire pour TIA@PFIA 2019 20
Vous pouvez aussi lire
