QASSIT: A Pretopological Framework for the Automatic Construction of Lexical Taxonomies from Raw Texts

Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 955–959, Denver, Colorado, June 4-5, 2015. c⃝2015 Association for Computational Linguistics QASSIT: A Pretopological Framework for the Automatic Construction of Lexical Taxonomies from Raw Texts Guillaume Cleuziou1, Davide Buscaldi2, Gael Dias3 Vincent Levorato4, Christine Largeron5 1 LIFO - University of Orl´eans, France 2 LIPN - University of Paris 13, France 3 GREYC - University of Caen-Basse Normandie, France 4 IRISE - CESI Orl´eans, France 5 LHC - University of Saint-Etienne, France cleuziou@univ-orleans.fr , davide.buscaldi@lipn.univ-paris13.fr gael.dias@unicaen.fr , vlevorato@cesi.fr christine.largeron@univ-st-etienne.fr Abstract This paper presents our participation to the SemEval Task-17, related to “Taxonomy Ex- traction Evaluation” (Bordea et al., 2015). We propose a new methodology for semi- supervised and auto-supervised acquisition of lexical taxonomies from raw texts. Our ap- proach is based on the theory of pretopology which offers a powerful formalism to model subsumption relations and transforms a list of terms into a structured term space by combin- ing different discriminant criteria. In order to reach a good pretopological space, we define the Learning Pretopological Spaces method that learns a parameterized space by using an evolutionary strategy. 1 Introduction Lexical Taxonomies (LTs) play an essential role in Information Retrieval (IR) and Natural Language Processing (NLP). By coding the semantic relations between terminological concepts, LTs can enrich the reasoning capabilities of applications in IR and NLP. However, the globalized development of semantic resources is largely limited by the efforts required for their construction (Kozareva and Hovy, 2010). As a consequence, instead of manually creating LTs, many research studies have emerged to automati- cally learn such structures (Buitelaar et al., 2005; Biemann, 2005; Cimiano et al., 2009; Kozareva and Hovy, 2010; Velardi et al., 2013). The two main stages for the automatic construc- tion of LTs are Term Extraction and Term Structur- ing. The proposed approach is focused on the sec- ond stage, thus matching with the aim of the Se- mEval task, by inducing LTs from pre-existing lists of terms (provided by the organizers). As starting point, we consider the work from (Cleuziou et al., 2011) which introduced new statistically-based criteria (e.g. Nearest-Neighbor- like relations) and combined them using the the- ory of pretopology (Brissaud, 1975). This formal- ism offers a new framework to model the subsump- tion relation at the term set level rather than consid- ering (binary) subsumption relations only between pairs of terms. Based on the concepts of (pseudo- )closure and closed subsets the authors transform the list of terms into a semantic space. A structur- ing algorithm based on the work of Largeron and Bonevay (2002) is then applied to transform the se- mantic space of terms into a LT i.e. an acyclic di- rected (non-triangular) graph. This theory should allow to combine both associative- and pattern-based methods within a vir- tuous multi-criteria structuring process. To achieve this objective, we consider pretopology on the multi- criteria analysis point of view, where criteria are statistical indices and linguistic patterns retrieved from a corpus. In particular, we define …

Content truncated to 500 words