Preliminary Experiments concerning Verbal Predicative Structure Extraction from a Large Finnish Corpus

The 3rd International Workshop for Computational Linguistics of Uralic Languages, pages 37–55, St. Petersburg, Russia, 23–24 January 2017. c⃝2017 Association for Computational Linguistics Preliminary Experiments in the Extraction of Predicative Structures from a Large Finnish Corpus Guersande Chaminade CNRS & ENS & PSL & Université Sorbonne nouvelle & USPC LATTICE Lab. & INaLCO 1 rue Maurice Arnoux 92120 Montrouge France gu.chaminade@gmail.com Thierry Poibeau CNRS & ENS & PSL & Université Sorbonne nouvelle & USPC LATTICE Lab. 1 rue Maurice Arnoux 92120 Montrouge France thierry.poibeau@ens.fr Abstract This paper presents different experiments aiming at automatically extracting lexical information from corpora. We first describe a simple experiment on the nature of direct objects in Finnish (partitive vs total objects) so as to check if data automatically extracted from corpora support traditional studies in the field. Which are the verbs subcategorizing partitive objects? Which are the verbs sub- categorizing (more frequently) total objects? We provide a tool and a user inter- face to browse data automatically collected from a large corpus in order to make it possible to answer this question. We then describe some ongoing work aiming at adapting to Finnish a complex lexical acquisition system initially developed for Japanese [1]. In our approach, following up on studies in natural language processing and linguistics, we embrace the double hypothesis: i) of a continuum between ambiguity and vagueness, and ii) of a continuum between arguments and adjuncts. We describe the system developed for the task and conclude with some perspectives for the evaluation of the produced resource. 1 37 1 Introduction Natural language processing traditionally requires precise and high coverage resour- ces to provide satisfactory results but it is well known that the development of such resources is long and costly. Recently machine learning has made it possible to de- velop resources at a lower cost. It is now possible to automatically analyse large corpora, typically made of several million words, and develop parsers based on the observation of surface regularities at corpus level, along with some annotated data used for training. Powerful unlexicalized parsers have been developed for several languages, with a surprisingly high accuracy given the fact no lexical information is provided in input [2, 3]. The output of these parsers has subsequently been used as a new source of knowl- edge for the development of large-scale lexical resources. This area of research, known as lexical acquisition, have permitted the development of large-scale dictionaries for example for English [4], French [5], Japanese [1] and lots of other languages as well. It has also been shown that the approach provides interesting (although not perfect) results: thanks to this approach, it for example possible to discover new subcatego- rization frames for particular verbs [6, 7] and to monitor in real time the evolution of word usage or the creation of new words in a language [1], etc. Results obtained with automatic methods are of course still far from perfect: they need to be manually checked but they usually provide lots of new results and …

Content truncated to 500 words