Is Knowledge-Free Induction of Multiword Unit Dictionary Headwords a Solved Problem?

Is Knowledge-Free Induction of Multiword Unit Dictionary Headwords a Solved Problem? Patrick Schone and Daniel Jurafsky University of Colorado, Boulder CO 80309 {schone, jurafsky}@cs.colorado.edu Abstract We seek a knowledge-free method for inducing multiword units from text corpora for use as machine-readable dictionary headwords. We provide two major evaluations of nine existing collocation-finders and illustrate the continuing need for improvement. We use Latent Semantic Analysis to make modest gains in performance, but we show the significant challenges encountered in trying this approach. 1 Introduction A multiword unit (MWU) is a connected collocation: a sequence of neighboring words “whose exact and unambiguous meaning or connotation cannot be derived from the meaning or connotation of its components” (Choueka, 1988). In other words, MWUs are typically non-compositional at some linguistic level. For example, phonological non-compositionality has been observed (Finke & Weibel, 1997; Gregory, et al, 1999) where words like “got” [g<t] and “to” [tu] change phonetically to “gotta” [g<rF] when combined. We have interest in inducing headwords for machine-readable dictionaries (MRDs), so our interest is in semantic rather than phonological non-compositionality. As an example of semantic non-compositionality, consider “compact disk”: one could not deduce that it was a music medium by only considering the semantics of “compact” and “disk.” MWUs may also be non-substitutable and/or non-modifiable (Manning and Schütze, 1999). Non- substitutability implies that substituting a word of the MWU with its synonym should no longer convey the same original content: “compact disk” does not readily imply “densely-packed disk.” Non- modifiability, on the other hand, suggests one cannot modify the MWU’s structure and still convey the same content: “compact disk” does not signify “disk that is compact.” MWU dictionary headwords generally satisfy at least one of these constraints. For example, a compositional phrase would typically be excluded from a hard-copy dictionary since its constituent words would already be listed. These strategies allow hard-copy dictionaries to remain compact. for As mentioned, we wish to find MWU headwords machine-readable dictionaries (MRDs). Although space is not an issue in MRDs, we desire to follow the lexicographic practice of reducing redundancy. As Sproat indicated, "simply expanding the dictionary to encompass every word one is ever likely to encounter is wrong: it fails to take advantage of regularities" (1992, p. xiii). Our goal is to identify an automatic, knowledge-free algorithm that finds all and only those collocations where it is necessary to supply a definition. “Knowledge-free” means that the process should proceed without human input (other than, perhaps, indicating whitespace and punctuation). This seems like a solved problem. Many collocation-finders exist, so one might suspect that most could suffice for finding MWU dictionary headwords. To verify this, we evaluate nine existing collocation-finders to see which best identifies valid headwords. We evaluate using two completely separate gold standards: (1) WordNet and (2) a compendium of Internet dictionaries. Although web-based resources are dynamic and have better coverage than WordNet (especially for acronyms and names), we show that WordNet-based scores are comparable to those using Internet MRDs. Yet the evaluations indicate that …

Content truncated to 500 words