Unsupervised Discovery of Scenario-Level Patterns for Information Extraction

Unsupervised Discovery of Scenario-Level Patterns for Information Extraction Roman Yangarber Ralph Grishman roman©as, nyu. edu grishman©cs, nyu. edu Courant Institute of Courant Institute of Mathematical Sciences Mathematical Sciences New York University New York University Pasi Tapanainen )~ Silja Huttunen $ tapanain©conexor, fi sihuttun~ling.helsinki, fi t Conexor Oy :~ University of Helsinki Helsinki, Finland Finland Abstract Information Extraction (IE) systems are com- monly based on pattern matching. Adapting an IE system to a new scenario entails the construction of a new pattern base---a time- consuming and expensive process. We have implemented a system for finding patterns au- tomatically from un-annotated text. Starting with a small initial set of seed patterns proposed by the user, the system applies an incremental discovery procedure to identify new patterns. We present experiments with evaluations which show that the resulting patterns exhibit high precision and recall. 0 Introduction The task of Information Extraction (I-E) is the selective extraction of meaning from free natural language text. I "Meaning" is under- stood here in terms of a fixed set of semantic objects--entities, relationships among entities, and events in which entities participate. The semantic objects belong to a small number of types, all having fixed regular structure, within a fixed and closely circumscribed subject do- main. The extracted objects are then stored in a relational database. In this paper, we use the nomenclature accepted in current IE literature; the term subject domain denotes a class of tex- tual documents to be processed, e.g., "business news," and scenario denotes the specific topic of interest within the domain, i.e., the set of facts to be extracted. One example of a sce- nario is "management succession," the topic of MUC-6 (the Sixth Message Understanding Con- ference); in this scenario the system seeks to identify events in which corporate managers left 1For general references on IE, cf., e.g., (Pazienza, 1997; muc, 1995; muc, 1993). their posts or assumed new ones. We will con- sider this scenario in detail in a later section describing experiments. IE systems today are commonly based on pat- tern matching. The patterns are regular ex- pressions, stored in a "pattern base" containing a general-purpose component and a substantial domain- and scenario-specific component. Portability and performance are two major problem areas which are recognized as imped- ing widespread use of IE. This paper presents a novel approach, which addresses both of these problems by automatically discovering good patterns for a new scenario. The viability of our approach is tested and evaluated with an actual IE system. In the next section we describe the problem in more detail in the context of our IE system; sec- tions 2 and 3 describe our algorithm for pattern discovery; section 4 describes our experimental results, followed by comparison with prior work and discussion, in section 5. 1 The IE System Our IE system, among others, contains a a back- end core engine, at the heart of which is a regular-e~xpression pattern matcher. The engine draws on attendant knowledge bases (KBs) of varying degrees of …

Content truncated to 500 words