Automatic Refinement of a POS Tagger Using a Reliable Parser and Plain Text Corpora

Automatic Refinement of a POS Tagger Using a Reliable Parser and Plain Text Corpora Hideki Hirakawa, Kenji Ono, Yumiko Yoshimura Human Interface Laboratory Corporate Research & Development Center Toshiba Corporation Komukai-Toshiba-cho 1, Saiwai-ku, Kawasaki, 212-8582, Japan {hideki.hirakawa, kenji2.ono, yumiko.yoshimura}@toshiba.co.jp Abstract This paper proposes a new unsupervised learning method for obtaining English part-of- speech(POS) disambiguation rules which would improve the accuracy of a POS tagger. This method has been implemented in the experimental system APRAS (Automatic POS Rule Acquisition System), which extracts POS disambiguation rules from plain text corpora by utilizing different types of coded linguistic knowledge, i.e., POS tagging rules and syntactic parsing rules, which are already stored in a fully implemented MT system. In our experiment, the obtained rules were applied to 1.7% of the sentences in a non-training corpus. For this group of sentences, 78.4% of the changes made in tagging results were an improvement. We also saw a 15.5 % improvement in tagging and parsing speed and an 8.0 % increase of parsable sentences. 1 Introduction Much research has been done on knowledge acquisition from large-scale annotated corpora as a rich source of linguistic knowledge. Major works done to create English POS taggers (henceforth, "taggers"), for example, include (Church 1988), (Kupiec 1992), (Brill 1992) and (Voutilainen et al. 1992). The problem with this framework, however, is that such reliable corpora are hardly available due to a huge amount of the labor-intensive work required. In case of the acquisition of non-core knowledge, such as specific, lexically or domain dependent knowledge, preparation of annotated corpora becomes more serious problem. One viable approach then is to utilize plain text corpora instead, as in (Mikheev 1996). But The method proposed by (Mikheev 1996) has its own weaknesses, in that it is restricted in scope. That is, it aims to acquire rules for unknown words in corpora from their ending characters without looking at the context. In the meantime, (Brill 1995a) (Brill 1995b) proposed a method to acquire context-dependent POS disambiguation rules and created an accurate tagger, even from a very small annotated text by combining supervised and unsupervised learning. The weakness of his method is that the effect of unsupervised learning decreases as the training corpus size increases. The problem in using plain text corpora for knowledge acquisition is that we need a human supervisor who can evaluate and sift the obtained knowledge. An alternative to this would be to use a number of modules of a well- developed NLP system which stores most of the highly reliable general rules. Here, one module functions as a supervisor for other modules, since all these modules are designed to work cooperatively and the knowledges stored in each module are correlated. Keeping this idea in mind, we propose a new unsupervised learning method for obtaining linguistic rules from plain text corpora using the existing linguistic knowledge. This method has been implemented in the rule extraction system APRAS (Automatic POS Rule Acquisition System), which automatically acquires rules for refining the morphological analyzer (tagger) in our …

Content truncated to 500 words