A New Approach to Text Understanding

A NEW APPROACH TO TEXT UNDERSTANDING Ralph Weischedel, Damaris Ayuso, Sean Boisen, Heidi Fox, Robert Ingria BBN Systems and Technologies 10 Moulton St. Cambridge, MA 02138 ABSTRACT This paper first briefly describes the architecture of PLUM, BBN's text processing system, and then reports on some experiments evaluating the effectiveness of the design at the component level. Three features are unusual in PLUM's architecture: a domain- independent deterrninistie parser, processing of (the resulting) fragments at the semantic and discourse level, and probabilistie models. 1. INTRODUCTION The PLUM (Probabilistic Language Understanding Model) natural language understanding system for extracting data from text is based on three unusual features: probabilistic language models, a domain-independent deterministic parser, and processing of (the resulting) fragments at the semantic and discourse level. Earlier papers have focused on the probabilistic aspects of the system [Weischedel et al., 1991; de Marcken, 1990]; here we focus on the other two design features. While several deterministic parsers have been constructed based on Marcus's Determinism Hypothesis, PLUM seems to be the first application system that employs a deterministic parser. Many systems have been built based on semantic and discourse-level processing of fragments, most notably systems based on conceptual dependency and scripts [Schank and Riesbeck, 1981]. However, PLUM may be the first system that uses a hybnd of such semantic techniques with the purely syntactic processing of Marcus's Determinism Hypothesis, two approaches that seemed totally antithetical when first proposed. The impact of marrying those two techniques is a robust system that produces answers in the application domain in spite of syntactic complexity, syntactic ill-formedness, extra-grammaticality and a high percentage of unknown words. A second impact is that the system can produce answers at a very early stage of porting it to a new domain. Both of these claims are substantiated in this paper by evaluating the performance of the system as the lexicon grows, without changing the syntactic, semantic, and discourse rules of the system. 2. BRIEF SYNOPSIS OF SYSTEM COMPONENTS Major system components are shown in the diagram in Figure 1. We expect the particular implementations to change and improve substantially during the next two years of research and development. A preprocessor driven by finite state rules divides the message into header material (if any), paragraphs, sentences, and trailer material (if any). A well-known problem in using deterministic parsing is the fact that most words in English are ambiguous even regarding part of speech. In the Foreign Broadcast Information Service texts of MUC-3, we estimate that the vocabulary had an average ambiguity of over two parts of speech in the TREEBANK tag system. age Morphological Analyzer Fast Partial Parser sem I I' Semantic Interpreter I I I • m o Fragment Combiner se[nn,, I I [ I Discourse Module s•m'n,,, I I eventl. ~ventm ~remplate Filler I I template 1 . .. template m, Figure 1. PLUM System Architecture In PLUM, determining the part of speech of highly ambiguous words is performed by well-known Markov modelling techniques. Though part of speech ambiguity was …

Content truncated to 500 words