Tagging for Learning: Collecting Thematic Relations from Corpus

Tagging for Learning: Collecting Thematic Relations from Corpus Uri Zernik and Paul Jacobs Artificial Intelligence Program GE Research and Development Center Schenectady, NY 12301 USA Abstract Recent work in text analysis has suggested that data on words that frequently occur to- gether reveal important information about text content. Co-occurrence relations can serve two main purposes in language pro- cessing. First, the statistics of co-occurrence have been shown to produce accurate results in syntactic analysis. Second, the way that words appear together can help in assign- ing thematic roles in semantic interpreta- tion. This paper discusses a method for col- lecting co-occurrence data, ~quiring lexical relations from the data, and applying these relations to semantic analysis. 1 Introduction Two text processing problems rely heavily on co- occurrence patterns- the way that words appear to- gether, possibly idiosyncraticly. First, statistically weighted co-occurrence information can assist in the "bracketing" of noun groups, which can otherwise lead to a eombinatoric explosion of parse trees [1]. Second, co-occurrence relations can provide evidence of semantic information for thematic-role assignment, an important task that is otherwise fraught with in- accuracy. Only co-occurrence patterns collected over a corpus can help to determine which is .object and which is recipient in PAID DIVIDEND (IS SECURE) vs. PAID SHAREHOLDERS (ARE SATISFIED). A sufficiently rich lexicon would include the semantic preferences for distinguishing these thematic roles, but such a lexicon does not yet exist. Co-occurrence patterns are a means of probing a global corpus for clues that help resolve ambiguity at the local sentence level. Patterns such as PAID TO SHAREHOLDERS and PAID THEM THE DIVIDEND are detected in the corpus at large. Through these latter examples, in which the distinction between re- cipient and object relative to the dative verb PAY is made explicit, the former cases in which tile relation is implicit can be resolved. In contrast to previous work which addressed the identification of surface relations, i.e., SVO triples [2], in our work we address the acquisition of semantic re- lations, focussing at the assigment of thematic roles. This task (i.e. tagging for acquisition) requires high reliability and so it relies less on statistical properties and more on deterministic local marking. In this paper we discuss a technique for parsing and semanticly analyzing complex sentences with the aid of co-occurrence relations, and show how these relations are acquired from tagged corpus. 1.1 The Phenomenon Consider, for example, the sentence below, taken from the Dow-Jones newswire: THE LARGEST CO~iPANY ON THE LIST, WHICH LAST PAID SHAREHOLDERS IN JANUARY, SAID THE 5 PC STOCK DIVIDEND WOULD BE PAYABLE FOLLOWING THE PAYMENT OF THE CASH DIVIDEND. (DJ, October 27, 1988) For this sentence, which is not exotic or unusual in its complexity, there are 24 non-trivial different parse trees. Human readers, in contrast to most programs, can quickly identify groups of words that "hang together" such as COMPANY PAID A DIVI- DEND, STOCK DIVIDEND, and CASH DIVIDEND, and use these clusters to understand the sentence unambiguously. Moreover, a human reader can …

Content truncated to 500 words