Automatic Term Extraction Based on Perplexity of Compound Words

Automatic Term Extraction Based on Perplexity of Compound Words Minoru Yoshida1,2 and Hiroshi Nakagawa1,2 1 Information Technology Center, University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033 2 JST CREST, Honcho 4-1-8, Kawaguchi-shi, Saitama 332-0012 mino@r.dl.itc.u-tokyo.ac.jp, nakagawa@dl.itc.u-tokyo.ac.jp Abstract. Many methods of term extraction have been discussed in terms of their accuracy on huge corpora. However, when we try to apply various methods that derive from frequency to a small corpus, we may not be able to achieve sufficient accuracy because of the shortage of statistical information on frequency. This paper reports a new way of extracting terms that is tuned for a very small corpus. It focuses on the structure of compound terms and calculates perplexity on the term unit’s left-side and right-side. The results of our experiments revealed that the accuracy with the proposed method was not that advantageous. However, experimentation with the method combining perplexity and frequency information obtained the highest average-precision in comparison with other methods. 1 Introduction Term extraction, which is the task of extracting terminology (or technical terms) from a set of documents, is one of major topics in natural language processing. It has a wide variety of applications including book indexing, dictionary generation, and keyword extraction for information retrieval systems. Most automatic term extraction systems make a sorted list of candidate terms extracted from a given corpus according to the “importance” scores of the terms, so they require scores of “importance” for the terms. Existing scores include TF-IDF, C-Value [1], and FLR [9]. In this paper, we propose a new method that involves revising the definition of the FLR method in a more sophisticated way. One of the advantages of the FLR method is its size-robustness, i.e, it can be applied to small corpus with less significant drop in performance than other standard methods like TF and IDF, because it is defined using more fine-grained features called term units. Our new method, called FPP, inherit this property while exhibiting better performance than FLR. At the same time, we also propose a new scheme for evaluating term ex- traction systems. Our idea is to use summaries1 of articles as a gold standard. This strategy is based on the assumption that summaries of documents can 1 In more detail, an article revised for display on mobile phones. R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 269–279, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 270 M. Yoshida and H. Nakagawa serve as collections of important terms because, in writing summaries, peo- ple may make an original document shorter by dropping unnecessary parts of original documents, while retaining essential fragments. Thus, we regard a term in an original document to be important if it also appears in the summary. 2 Term Extraction Term extraction is the task of extracting important terms from a given corpus. Typically, term extraction systems first extract term candidates, which are usu- ally the noun phrases detected by handcrafted POS sequence patterns, from the corpus. After that, term candidates are sorted according to …

Content truncated to 500 words