Towards Terascale Semantic Acquisition

Towards Terascale Knowledge Acquisition Patrick Pantel, Deepak Ravichandran and Eduard Hovy Information Sciences Institute University of Southern California 4676 Admiralty Way Marina del Rey, CA 90292 {pantel,ravichan,hovy}@isi.edu Abstract Although vast amounts of textual data are freely available, many NLP algorithms exploit only a minute percentage of it. In this paper, we study the challenges of working at the terascale. We present an algorithm, designed for the terascale, for mining is-a relations that achieves similar performance to a state-of-the-art linguistically-rich method. We fo- cus on the accuracy of these two systems as a func- tion of processing time and corpus size. 1 Introduction The Natural Language Processing (NLP) com- munity has recently seen a growth in corpus-based methods. Algorithms light in linguistic theories but rich in available training data have been success- fully applied to several applications such as ma- chine translation (Och and Ney 2002), information extraction (Etzioni et al. 2004), and question an- swering (Brill et al. 2001). In the last decade, we have seen an explosion in the amount of available digital text resources. It is estimated that the Internet contains hundreds of terabytes of text data, most of which is in an unstructured format. Yet, many NLP algorithms tap into only megabytes or gigabytes of this information. In this paper, we make a step towards acquiring semantic knowledge from terabytes of data. We present an algorithm for extracting is-a relations, designed for the terascale, and compare it to a state of the art method that employs deep analysis of text (Pantel and Ravichandran 2004). We show that by simply utilizing more data on this task, we can achieve similar performance to a linguistically- rich approach. The current state of the art co- occurrence model requires an estimated 10 years just to parse a 1TB corpus (see Table 1). Instead of using a syntactically motivated co-occurrence ap- proach as above, our system uses lexico-syntactic rules. In particular, it finds lexico-POS patterns by making modifications to the basic edit distance algorithm. Once these patterns have been learnt, the algorithm for finding new is-a relations runs in O(n), where n is the number of sentences. In semantic hierarchies such as WordNet (Miller 1990), an is-a relation between two words x and y represents a subordinate relationship (i.e. x is more specific than y). Many algorithms have recently been proposed to automatically mine is-a (hypo- nym/hypernym) relations between words. Here, we focus on is-a relations that are characterized by the questions “What/Who is X?” For example, Table 2 shows a sample of 10 is-a relations discovered by the algorithms presented in this paper. In this table, we call azalea, tiramisu, and Winona Ryder in- stances of the respective concepts flower, dessert and actress. These kinds of is-a relations would be useful for various purposes such as ontology con- struction, semantic information retrieval, question answering, etc. The main contribution of this paper is a compari- son of the quality of our pattern-based and co- occurrence models as a function of processing time …

Content truncated to 500 words