Automatic Learning for Semantic Collocation

Automatic Learning for Semantic Collocation Satoshl SEKINE* Tokyo Information and Communications Research Laboratory Matsushita Electric Industrial Co.,Ltd. 3-10-1, higashimita, tama-ku, kawasaki 214 JAPAN Jeremy J.CARROLL Sofia ANANIADOU Jun'ichi TSUJII Centre for Computational Linguistics University of Manchester Institute of Science and Technology P.O.Box 88, Manchester M60 1QD, United Kingdom Abstract The real difficulty in development of practical NLP systems comes from the fact that we do not have effective means for gathering "knowl- edge". In this paper, we propose an algorithm which acquires automatically knowledge of se- mantic collocations among "words" from sam- ple corpora. The algorithm proposed in this paper tries to discover semantic collocations which will be useful for disambiguating structurally ambigu- ous sentences, by a statistical approach. The algorithm requires a corpus and minimum lin- guistic knowledge (parts-of-speech of words, simple inflection rules, and a small number of general syntactic rules). We conducted two experiments of applying the algorithm to different corpora to extract dif- ferent types of semantic collocations. Though there are some unsolved problems, the results showed the effectiveness of the proposed algo- rithm. 1 Introduction Quite a few grammatical formalisms have been proposed by computational linguists, which are claimed to be "good" (declarative, highly modular, etc.) for practi- cal application systems in NLP. It has also been claimed that extra-linguistic, domain specific knowledge is in- dispensable in most NLP applications, and computa- tional frameworks for representing and using such do- main knowledge have also been developed. However, the real difficulty in developing practical NLP systems is due to the fact that we do not have effec- tive means for gathering the "knowledge", whether lin- *SEKINE is now a visitor at C.C.L., U.M.I.S.T. s ekine @ ccl. umist, ac. uk guistic or extra-linguistic. In particular, it has been re- ported [Ananiadou, 1990] that not only extra-linguistic, domain knowledge but also linguistic knowledge required for application systems varies, depending on text-type (technical reports, scientific papers, manuals, etc.), sub- ject domain, type of application (MT, automatic ab- straction, etc.) etc. This means that we have to have ef- fective and efficient methods either for adapting already existing knowledge for a specific "sublanguage" or for ac- quiring knowledge automatically, for example from sam- ple corpora of given applications. In this paper, we propose an algorithm which auto- matically acquires knowledge of semantic collocations among "words". "Semantic" here means that the col- locations the algorithm discovers are not collocations among words in the sense of traditional linguistics but collocations that reflect ontological relations among en- tities in given subject domains. We expect that the knowledge to be extracted will not only be useful for disambiguating sentences but also will contribute to dis- covering ontological classes in given subject domains. Though several studies with similar objectives have been reported [Church, 1988], [Zernik and Jacobs, 1990], [Calzolari and Bindi, 1990], [Garside and Leech, 1985], [Hindle and Rooth, 1991], [Brown et al., 1990], they require that sample corpora be correctly analyzed or tagged in advance. It must be a training corpus, which is …

Content truncated to 500 words