In the literature on lexical acquisition and word sense disambiguation, many doubts are expressed about the usefulness of MRDs and corpora for full-blown acquisition of a computational lexicon. I present a selection of them here, focusing on those which have to do with sense discrimination.
atkins:91 asks some fundamental questions about the nature of dictionaries, in comparison to the needs of computational lexica.
[H]ow much semantic information accurate enough to be useful in a computational lexicon is contained in a dictionary definition written for the human user, who often consciously supplements and corrects what is being read? Is it indeed possible to write dictionary definitions that encapsulate the essentail facts about the senses of a word? Can the meaning of a word be divided into discrete senses without distorting it beyond reason? (Atkins 1991:168)The conclusion from her investigations is that objective evidence supporting the accuracy of dictionary entries does not exist; that there is little systematicity in the sense differentiations made for a word across dictionaries. She argues:
The traditional dictionary entry is trying to do what the language simply will not allow. Word meaning cannot be sliced up into distinct bundles, labelled (however carefully) and packaged into a dictionary entry which will tell the truth, the whole truth and nothing but the truth. (Atkins 1991:180)atkins:91 and atkins_levin:91 suggest that MRDs can only be effectively utilised in conjunction with a comprehensive theory of the lexicon, which defines `templates' for lexical entries, establishes what information about particular words might be expected to be found in the dictionary entries for the word, given the identification of the word's semantic class. This approach, then, requires an interaction between the information in the MRD and the structure of the lexicon set out by the theory.
The need for a theory of the lexicon is also implicit in the following statement by Ann Copestake:
The process of constructing lexical entries [from LDOCE] results in [a Lexical Knowledge Base] which is defective in that it retains the LDOCE sense distinctions. [...] There is little alternative to this currently, given the lack of an adequate theory of sense distinction, although it would clearly be desirable to have a more linguisticaly motivated treatment. (Copestake 1992:136)
The survey of MRD research presented above (Section 6.5.1) showed how narrow the range of semantic information successfully extracted from MRDs to date has been. There are many more relationships among words and word senses which are important for NLP applications than the hyperonymy relation. Synonymy/antonymy and meronymy/holonymy, for example, are important structuring relations (Miller miller:90), as is knowledge of regular polysemy. The work of davis:95, as discussed in Chapter 2, also strongly indicates that semantic relations at a level deeper than semantic class, the level of semantic roles, is critical for capturing generalisations about language structure. Whether such information exists within MRDs, even implicitly, and whether it could be extracted automatically in a consistent way is not immediately obvious.
For corpus-based lexicon acquisition, the picture is even more bleak. There is no structure in corpora whatsoever, and statistical techniques can only provide very coarse distinctions in word use. The linguistic information necessary for identification of regular polysemies and for accurate classification of words into taxonomies is simply not available from a corpus. Corpora are best used as a data resource for evidence of how particular words are used, and for identification of collocations and non-lexical units (e.g. idioms), in conjunction with analytical methodology for identifying relationships between uses of words.
The fundamental problem for these automatic techniques is that they depend on pre-existing divisions between word senses which, as we saw in Section 6.3, are not easily justifiable and cannot by their nature be fully identified in isolation of particular contexts. These approaches seemingly deny the creative aspect of language use from the outset and will therefore always fall short of the ultimate goal of identifying the underlying principles of generative language use. NLP systems which require sophisticated language processing demand a framework which will accommodate the flexibility of language use and which will result in fine-grained interpretation. This framework can only come from linguistic theory.