Fully Unsupervised Graph-Based Discovery of General-Specific Noun Relationships from Web Corpora Frequency Counts

CoNLL 2008: Proceedings of the 12th Conference on Computational Natural Language Learning, pages 97–104 Manchester, August 2008 Fully Unsupervised Graph-Based Discovery of General-Specific Noun Relationships from Web Corpora Frequency Counts Gaël Dias HULTIG University of Beira Interior ddg@di.ubi.pt Raycho Mukelov HULTIG University of Beira Interior raicho@hultig.di.ubi.pt Guillaume Cleuziou LIFO University of Orléans cleuziou@univ-orleans.pt Abstract. In this paper, we propose a new metho- dology based on directed graphs and the TextRank algorithm to automatically in- duce general-specific noun relations from web corpora frequency counts. Different asymmetric association measures are im- plemented to build the graphs upon which the TextRank algorithm is applied and produces an ordered list of nouns from the most general to the most specif- ic. Experiments are conducted based on the WordNet noun hierarchy and assess 65.69% of correct word ordering. 1 Introduction Taxonomies are crucial for any knowledge- based system. They are in fact important because they allow to structure information, thus foster- ing their search and reuse. However, it is well known that any knowledge-based system suffers from the so-called knowledge acquisition bottle- neck, i.e. the difficulty to actually model the do- main in question. As stated in (Caraballo, 1999), WordNet has been an important lexical know- ledge base, but it is insufficient for domain spe- cific texts. So, many attempts have been made to automatically produce taxonomies (Grefenstette, 1994), but (Caraballo, 1999) is certainly the first work which proposes a complete overview of the problem by (1) automatically building a hierar- chical structure of nouns based on bottom-up clustering methods and (2) labeling the internal nodes of the resulting tree with hypernyms from the nouns clustered underneath by using patterns such as “B is a kind of A”. © 2008. Licensed under the Creative Commons At- tribution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc- sa/3.0/). Some rights reserved. In this paper, we are interested in dealing with the second problem of the construction of an or- ganized lexical resource i.e. discovering general- specific noun relationships, so that correct nouns are chosen to label internal nodes of any hierar- chical knowledge base, such as the one proposed in (Dias et al., 2006). Most of the works pro- posed so far have (1) used predefined patterns or (2) automatically learned these patterns to identi- fy hypernym/hyponym relationships. From the first paradigm, (Hearst, 1992) first identifies a set of lexico-syntactic patterns that are easily recog- nizable i.e. occur frequently and across text genre boundaries. These can be called seed patterns. Based on these seeds, she proposes a bootstrap- ping algorithm to semi-automatically acquire new more specific patterns. Similarly, (Carabal- lo, 1999) uses predefined patterns such as “X is a kind of Y” or “X, Y, and other Zs” to identify hypernym/hyponym relationships. This approach to information extraction is based on a technique called selective concept extraction as defined by (Riloff, 1993). Selective concept extraction is a form of text skimming that selectively processes relevant text while effectively ignoring surround- ing text that is thought to be irrelevant to the do- …

Content truncated to 500 words