Superior and Efficient Fully Unsupervised Pattern-based Concept Acquisition Using an Unsupervised Parser

Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL), pages 48–56, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Superior and Efficient Fully Unsupervised Pattern-based Concept Acquisition Using an Unsupervised Parser Dmitry Davidov1 Roi Reichart1 Ari Rappoport2 1ICNC , 2Institute of Computer Science Hebrew University of Jerusalem {dmitry@alice.nc|roiri@cs|arir@cs}.huji.ac.il Abstract Sets of lexical items sharing a significant aspect of their meaning (concepts) are fun- damental for linguistics and NLP. Unsuper- vised concept acquisition algorithms have been shown to produce good results, and are preferable over manual preparation of con- cept resources, which is labor intensive, er- ror prone and somewhat arbitrary. Some ex- isting concept mining methods utilize super- vised language-specific modules such as POS taggers and computationally intensive parsers. In this paper we present an efficient fully unsupervised concept acquisition algorithm that uses syntactic information obtained from a fully unsupervised parser. Our algorithm incorporates the bracketings induced by the parser into the meta-patterns used by a sym- metric patterns and graph-based concept dis- covery algorithm. We evaluate our algorithm on very large corpora in English and Russian, using both human judgments and WordNet- based evaluation. Using similar settings as the leading fully unsupervised previous work, we show a significant improvement in con- cept quality and in the extraction of multiword expressions. Our method is the first to use fully unsupervised parsing for unsupervised concept discovery, and requires no language- specific tools or pattern/word seeds. 1 Introduction Comprehensive lexical resources for many domains and languages are essential for most NLP applica- tions. One of the most utilized types of such re- sources is a repository of concepts: sets of lexical items sharing a significant aspect of their meanings (e.g., types of food, tool names, etc). While handcrafted concept databases (e.g., Word- Net) are extensively used in NLP, manual compila- tion of such databases is labor intensive, error prone, and somewhat arbitrary. Hence, for many languages and domains great efforts have been made for au- tomated construction of such databases from avail- able corpora. While language-specific and domain- specific studies show significant success in develop- ment of concept discovery frameworks, the majority of domains and languages remain untreated. Hence there is a need for a framework that performs well for many diverse settings and is as unsupervised and language-independent as possible. Numerous methods have been proposed for seed- based concept extraction where a set of concept pat- terns (or rules), or a small set of seed words for each concept, is provided as input to the concept acqui- sition system. However, even simple definitions for concepts are not always available. To avoid requiring this type of input, a number of distributional and pattern-based methods have been proposed for fully unsupervised seed-less acquisi- tion of concepts from text. Pattern-based algorithms were shown to obtain high quality results while be- ing highly efficient in comparison to distributional methods. Such fully unsupervised methods do not incorporate any language-specific parsers or taggers, so can be successfully applied to diverse languages. However, unsupervised pattern-based …

Content truncated to 500 words