Class notes ECON 103

14.2 Partial parsing and LTAG A chunk based partial parser is necessitated from more requirements in un- restricted dialog systems. Due to inherent characteristics of speech, the grammatical coverage of these systems is seldom complete. In general full parsing requires, enormous amount of semantic knowledge (unrestricted do- main dialogs are often ungrammatical and semantic knowledge sources are not sufficiently viable for guiding the parser). An other advantage of the chunk based parsing is that chunks can be identified with very simple finite-state devices. The possibility of handling the two main tasks of the parser separately (the identification and the at- tachment of chunks) is the principal reason for this simplification ([1]). Based on psycholinguistic studies of ([7]), ([1]) define the chunk as a syntactic structure, which comprises a connected subgraph of the sentence’s parse tree (..). These chunks are defined in terms of major heads. It should be noted that chunks do not necessarily partition the sentence. There may be material that does not belong to any chunk. For example, prepositions, coordinators, subordinators, adverbs and punctuation often are not part of any chunk ([3]). The major problem of this approach is that the semantic interpretation is limited only on chunks with verbal heads ([22]). The direct effect of this limitation is that the semantic interpretation corresponding to syntactic parsing is very poor and practically useless (the adverb, the coordination and the preposition are essential for the construction of the interpretation). Thus following [([20]), ([19])] we decide to use the Lexicalized Tree Ad- joining Grammar formalism LTAG [for an introduction to this formalism, see ([10])]. For many reasons, this formalism seems to be adapted for lan- guage modeling and the chunk based partial parsing approach. First, el- ementary trees compile out the various syntactic environments the anchor of the tree can appear in. Second, elementary trees may be regarded as linguistically motivated chunk templates with slots for the arguments of the anchor. Third, since every tree is lexicalized in LTAG, words are associated more tightly with the structures they anchor. This allows the possibility of making the operations of combining elementary trees to be sensitive to lexical context. ...........................................CONTINUED.................................