Arguments for Parallel Distributed Parsing: Toward the Integration of Lexical and Sublexical (Semantic) Parsings

Arguments for Parallel Distributed Parsing Toward the integration of lexical and sublexical (semantic) parsings Kow KURODA National Institute of Information and Communication Technologies (NICT), Japan Abstract. This paper illustrates the idea of parallel distributed parsing (PDP), which allows us to integrate lexical and sublexical analyses. PDP is proposed for providing a new model of efficient, information-rich parses that can remedy the data sparseness problem. 1 Introduction In the usual sense of syntactic parsing, the analyses of relationships among words and relationships among morphemes are separated. The former is called syntactic analysis (or syntactic parsing) and the latter is called morphological analysis (the term “morphological parsing” exists but it seems to denote a somewhat different notion). Sometimes, however, sublexical analysis is relevant. This is evident in shallow semantic parsing which is understood to be “labeling phrases of a sentence with semantic roles with respect to a target word. For example, the sentence (1) is labeled as (2):”1) (1) Shaw Publishing offered Mr. Smith a reimbursement last March. (2) [Agent Shaw Publishing ] offered [Recipient Mr. Smith ] [Theme a reimbursement ] [Time last March ] . The target of the labeling in (2) is offered. Note that the same kind of labeling should be available for the argument structure of reimbursement, which can be illustrated in (3): (3) [Payer(asAgent) Shaw Publishing] offered [Recipient Mr. Smith ] a [Target reimburse]-ment last March. (No explicit mention of Payment (as Theme)) Clearly, semantic role labeling requires a high-precision predicate-argument analysis of a target predi- cate, whether the target is a word (e.g., offer) or a morpheme (e.g., reimburse) embedded in a word. The comparison of the two cases shows that a certain kind of parsing is necessary in effective semantic labeling, as Gildea and Palmer (2002) pointed out. But the problem is how to combine the lexical parse by which the predicate-argument structure of offer is recognized with the sublexical parse by which the predicate- argument structure of reimburse is recognized. Integrating the two kinds of parses is not a trivial task. This paper presents the idea of Parallel Distributed Parsing (PDP) that is able to straightforwardly carry out the integration. The presentation, however, is theoretically oriented and the content is rather preliminary: no empirical results are presented other than a few sample parses. No parser implementation is available. The main purpose of this presentation is to illustrate a new model for parsing that integrates lexical and sublexical parsings, which I argue can be a remedy for the problem of data sparseness. Data sparseness is a serious problem in natural language processing (NLP) even now that computers can access more raw data than the average human does. The size of textual raw data automatically acquired from the Web exceeds that which a normal human can read in a lifetime. This suggests, however, that no human seems to suffer from data sparseness. What makes this more mysterious is that humans do also employ statistical information in their language processing. The difference between humans and machines, …

Content truncated to 500 words