Inferences for Lexical Semantic Resource Building with Less Supervision

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 2300–2305 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 2300 Inferences for Lexical Semantic Resource Building with Less Supervision Nadia Bebeshina1,2, Mathieu Lafourcade2 1Praxiling, 2LIRMM Universit´e Paul Val´ery, Universit´e de Montpellier, France nadia.clairet@gmail.com, lafourcade@lirmm.fr Abstract Lexical semantic resources may be built using various approaches such as extraction from corpora, integration of the relevant pieces of knowledge from the pre-existing knowledge resources, and endogenous inference. Each of these techniques needs human supervision in order to deal with the potential errors, mapping difficulties or inferred candidate validation. We detail how various inference processes can be employed for lexical semantic resource building with less supervision. Our experience is based on the combination of different inference techniques for multilingual resource building and evaluation. Keywords: inference, multilingual knowledge resource, semantic relations, validation 1. Introduction The lexical and semantic resource building based on infer- ence represents an appealing area of research. Inference consists in proposing new elements automatically based on the existing ones. In the context of a lexical semantic re- source, inference is a process of calculating new pieces of knowledge i.e. semantic relations without using any ex- ternal structured knowledge repository. Thus, an inference engine can be seen as a system that, given a set of rules, provides the action of creating the new elements in the re- source. Thus, such this mechanism allows building lexi- cal semantic resource incrementally and, presumably, min- imise the human effort necessary for the resource building. Unfortunately, the knowledge parts obtained by inference have to be semi-automatically double checked to avoid the propagation of errors. The part of human effort in the infer- ence process may be very important to perform rule design, validation, resource improvement. In the present paper, we describe a resource building pipeline based on monolingual inference and cross-lingual inference. Its main contribution is proposing a method where the resource building process appears as a self learn- ing process as the feedback of the evaluation by inference impacts the resource that is being built. The paper is organised as follows. First, we introduce the state of the art techniques for lexical semantic resource building. Second, we detail the resources we use for the inference based resource building. Third, we provide the details about the resource building and evaluation experi- ence. 2. Related Work Cross-lingual relationship inference benefits from active re- search efforts. In the framework of the large knowledge bases (KBs) such as NELL (Mitchell et al., 2015), several approaches focused on the equivalence between entities and relationships have been introduced. For instance, authors in (Hern´andez-Gonz´alez et al., 2017) describe the experience of merging several monolingual editions of NELL. Authors in (Nickel et al., 2015) detail the statistical relational learn- ing on knowledge graphs (KGs) and point out the impor- tance of type constraints and transitivity. Similar to (Wang et al., 2015), they base their method mainly on large scale KBs such as Nell (Carlson et al., …

Content truncated to 500 words