acquire contcxt-dcpendent POS disambiguation rules and created an accurate tagger, even from a very small aunotated text by combining supervised and tmsupcrviscd learning. The wcakness 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 welldeveloped NLP system which stores most of thc highly reliable general rules. Here, one module fimctions as a supervisor for other modules, since all these modules arc designed to work cooperatively and the knowledgcs tored in each module are correlated. Keeping this idea in mind, we propose a new unsupervised learning method for obtaining linguistic rules fi'om 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 rnles for refining the morphological analyzer (taggcr) in our English-Japanese MT system ASTRANSAC (Hirakawa ct al. 1991) through the interaction between the system's tagger and parser on the assumption that they are considerably accurate. This paper is organized as follows: Section 2 illustrates the basic idea of our method; Section 3 gives the outline of APRAS; Sections 4 and 5 describe our experiments. 2 Basic Idea Our MT system has a tagger which can generate ranked POS sequences of input sentences according to their plausibility and also a parser which judges the parsability of the derived POS sequences one by one until a parsable one is found ~ . in our framework, this tagger can be viewed as a POS candidate generator, and the parser as a sifter. Now sentences can be categorized into the following three: (P) a balanced sentence, whose top ranked

sequence, or initial POS sequence, is

parsable, (Q) a conflicting sentence, in which the top

ranked scquencc is unparsable, but there are

parsable ones in the rest of the sequences;

and (R) an unparsable sentence, in which all the

POS sequences are unparsable. Before going on to our main discussion, we will briefly explain the terminology used in this paper. Here we call a highest-ranking parsable POS sequence as the "Most Preferable _PParsable POS sequence," or simply "MPP POS sequence." For our purposes, we will make use of balanced sentences and conflicting sentences. We call the POS of a word in the initial POS sequence as its "initially tagged POS" and that in the MPP POS sequence as its "parsable POS." We call the word whose initially tagged POS and parsable POS differ as a "focus word." Since the tagger is accurate, we can expect only a few POS differences between the initial and MPP POS sequences for a sentence. Finally, let us call

Here only top-N POS scquenccs are tried, where N is a pre-defined constant o limit parsing time. the POS's of the preceding and succeeding words as the "POS context of the focus word." Conflicting sentences, and their initial POS sequences, parsable POS sequences, and focus words can be automatically cxtracted. Through extraction out of a large amount of plain text corpora combined with statistical filtering, it would be possible to automatically select the proper POS conditions that could determine POS's of focus words. Then, we extract "POS Adjusting rules" or "PA rules" defined as below: PA rule: W(IPOS) ---> W(PPOS): C C: Context W :Word IPOS: Initially tagged POS PPOS: Parsable POS

Means "Give priority to the parsable POS

over the initially tagged POS in a particular

context shown as 'C'." PA rules do not determine POS's of words from their context, but change the judgement made by the tagger in a particular context. Extracted PA rules arc independent rules to the tagger and the parser used in the extraction. At the samc time, these rules are optimized for the tagger and the parser, since they are derived only from conflicting sentences, not from balanced sentences. Hence, the knowledge already coded in the system will not be extracted. I11 the following section, we give the outline of APRAS focusing on its two modules. 3 Outline of APRAS Fig. 1 shows the application of APRAS to an MT system. APRAS works in two phases, a rule extraction phase and a rule application phase. Note that the same tagger and the parser of the MT system are used throughout. I'A rule extraction phase PA rule application phase ---

qYaining Coqms Mf Systcm ~APP, AS M'I' Systm~ ~ Rule caMidatc

gctlcralion J \[ GenorationJ \[ Generation \] Figure 1 : Application of APRAS to an MT System In the rule extraction phase, the tagger analyzes each sentence in a training corpus and produces plausible POS sequences. The parser then judges the parsability of each POS sequence. Whenever a conflicting sentence appears, the role generation module outputs the candidates of PA rules. After all PA rule candidates for this training corpus are generated, the rule filtering module statistically weighs the validity of obtained PA rule candidates, and filters out unreliable rules. Sentences in the training corpus are not mmslated in this phase. In the rule application phase, both the already installed POS rulcs and the acquired PA rules are used for tagging. A sentence is parsed and then translated into target language. PA rules basically act to avoid the taggcr's wasteful generation of POS sequences. This would improve the ranking of POS sequences the tagger outputs and also increase the chances that the parser will find a parsable or better POS sequence in the improved ranking. 3.1 Rule Generation Module PA rule candidates are generated fronl conflicting sentences. Balanced and unparsable sentences generate no PA rule candidate. The words in balanced sentences arc recorded along with their POS's and POS contexts to be used in the rule filtering module. Whenever the system enconllters a conflicting sentence in a training corpus, the system compares the initial POS sequence with the MPP POS sequence of the scntcncc and picks up focus words. Then, for every focus word, the system generates a PA rule candidate which consists of a focus word, its initially tagged POS, parsable POS, and the POS context, i.e., the preceding POS's and the succeeding POS's. t:ig. 2 illustrates how a PA rulc candidate is generated. The focus word is 'rank', its initially tagged POS is "(verb)', its parsable POS is "(noun)', and the POS context is "(verb)(determiner)-$-'in'-(dcterminer)", where "$' denotes the focus word. The POS context is composed of preceding two POS's and succeeding two POS's. ttere surface words can be used instead of POS, like "in' in the example. The generated PA rule candidate can be read as: If the word 'rank' appears in a POS context "(verb)-(determiner)-$-'in'-(determiner)", then give priority to "(noun)' over "(verb)'. In this rule generation module, two important factors should be taken into account: namely, context size and levels of abstraction. If we expand the context of a focus word, the PA rule should gain accuracy. But its frequency in the training corpus would drop, thereby making it difficult to perform statistical filtering. To ensure statistical reliability, we need a largeInput sentence • • • n\]ove

(Focus word) 1 ,, ,, I the ', rank ', in I I I POS tagger output ,, ', I

,, the', ... I Parser output Initial POS sequence o ° ° . , ° MPP POS sequence PA rule candidate: Initially tagged POS Parsable POS POS context Figure 2: PA Rule Candidate Generation sized training corpus. At present we set the context size to be two words. In choosing adequate levels of abstraction or specification of POS in the context, we grouped together those POS tags which influence the choice of POS of a focus word in a similar lnanner as one super-POS tag, as in (Haruno & Matsumoto 1997). We also changed some POS tags for functional words like prepositions and words such as "be' and "have' to tags which denote their literal forms, because the choice of POS of a focus word is highly dependent on the word itsclf. As a result, we obtained 513 POS tags including 16 POS tags for nouns, 17 for verbs, 410 for prepositions and phrasal prepositions, and 70 for adjectives and adverbs.