A Memory-Based Approach to Learning Shallow Natural Language Patterns

A Memory-Based Approach to Learning Shallow Natural Language Patterns Shlomo Argamon and Ido Dagan and Yuval Krymolowski Department of Mathematics and Computer Science Bar-Ilan University 52900 Ramat Gan, Israel { argamon, dagan, yuvalk}@cs, biu. ac. il Abstract Recognizing shallow linguistic patterns, such as ba- sic syntactic relationships between words, is a com- mon task in applied natural language and text pro- cessing. The common practice for approaching this task is by tedious manual definition of possible pat- tern structures, often in the form of regular expres- sions or finite automata. This paper presents a novel memory-based learning method that recognizes shal- low patterns in new text based on a bracketed train- ing corpus. The training data are stored as-is, in efficient suffix-tree data structures. Generalization is performed on-line at recognition time by compar- ing subsequences of the new text to positive and negative evidence in the corpus. This way, no in- formation in the training is lost, as can happen in other learning systems that construct a single gen- eralized model at the time of training. The paper presents experimental results for recognizing noun phrase, subject-verb and verb-object patterns in En- glish. Since the learning approach enables easy port- ing to new domains, we plan to apply it to syntac- tic patterns in other languages and to sub-language patterns for information extraction. 1 Introduction Identifying local patterns of syntactic sequences and relationships is a fundamental task in natural lan- guage processing (NLP). Such patterns may corre- spond to syntactic phrases, like noun phrases, or to pairs of words that participate in a syntactic rela- tionship, like the heads of a verb-object relation. Such patterns have been found useful in various application areas, including information extraction, text summarization, and bilingual alignment. Syn- tactic patterns are useful also for many basic com- putational linguistic tasks, such as statistical word similarity and various disambiguation problems. One approach for detecting syntactic patterns is to obtain a full parse of a sentence and then extract the required patterns. However, obtaining a complete parse tree for a sentence is difficult in many cases, and may not be necessary at all for identifying most instances of local syntactic patterns. An alternative approach is to avoid the complex- ity of full parsing and instead to rely only on local information. A variety of methods have been devel- oped within this framework, known as shallow pars- ing, chunking, local parsing etc. (e.g., (Abney, 1991; Greffenstette, 1993)). These works have shown that it is possible to identify most instances of local syn- tactic patterns by rules that examine only the pat- tern itself and its nearby context. Often, the rules are applied to sentences that were tagged by part- of-speech (POS) and are phrased by some form of regular expressions or finite state automata. Manual writing of local syntactic rules has become a common practice for many applications. However, writing rules is often tedious and time consuming. Furthermore, extending the rules to different lan- guages or sub-language domains can require …

Content truncated to 500 words