Developing knowledge bases for MT with linguistically motivated quality-based learning

MT Summit VII Sept. 1999 Developing Knowledge Bases for MT with Linguistically Motivated Quality-Based Learning Evelyne Viegas New Mexico State University Computing Research Laboratory Las Cruces, NM 88003 USA viegas@crl.nmsu.edu Abstract In this paper we present a proposal to help bypass the bottleneck of knowledge-based sys- tems working under the assumption that the knowledge sources are complete. We show how to create, on the fly, new lexicon entries us- ing lexico-semantic rules and how to create new concepts for unknown words, investigating a new linguistically-motivated model to trigger concepts in context. 1 Introduction When parsing or generating, the MT sys- tem heavily depends on the completeness of the static sources available to it, in terms of Lexical Knowledge Base (LKB), Conceptual Knowledge Base (CKB) or ontology, and the grammar. Fighting the incompleteness of the static sources has been addressed by both the Computational Linguistic (CL) and the Arti- ficial Intelligence (AI) communities. CL re- searchers extracted knowledge (usually lexi- cal and/or syntactic) from very large corpora. This lexical acquisition has been mainly con- cerned with surface level co-occurring data (e.g., Zernik and Jacobs, 1990), hyponymy extraction based on syntactic criteria (e.g., Hearst, 1992) or lexico-semantic associations (e.g., Resnik, 1992). However these proposals cannot go be- yond a shallow level of learning (selectional re- strictions). Yet understanding needs a more fine-grained conceptual knowledge, in terms of roles and role filler constraints (e.g., Mahesh et al. (1997a) for a discussion on the neces- sity of such knowledge.) In this proposal, we only use positive occurrences of a small corpus of one text (there is however no limitation on the length of a text; in fact, the longer, the better) to avoid dealing with multiple mean- ings of a word used in different contexts (e.g. homonymy). AI researchers used standard ma- chine learning algorithms or adapted them to the language understanding task (see Wermter et al. (1996), for a review). As pointed out by Schnattinger and Hahn (1998), the main draw- back of these proposals is that they keep the un- derstanding and learning modes separate. How- ever the interpretation of a text depends on both, meaning that these modes have to interact with each other. In this proposal, we advocate a knowledge-intensive model of word and concept learning which interacts with the non-learning mode of text understanding. At the lexical level (Section 2), we make use of lexico-semantic rules (LSRs) to create new entries. LSRs create new syntactic and seman- tic mappings in a new entry. At the conceptual level (Section 3), we investigate a model to cre- ate new concepts for unknown words. To do so, we use the analysis of the text to provide the unknown word with its conceptual environ- ment in a given context, and a concept trigger model, we call Concept Trigger (CT), which builds the semantic space around the new con- cept (in other words, the conceptual frame) and attempts to place it in the CKB or ontology hi- erarchy. Briefly, when a word in a …

Content truncated to 500 words