Modelling syntactic uncertainty in lexical acquisition from texts*
Roberto Basili, Alessandro Marziali, Maria Teresa Pazienza · Journal of Quantitative Linguistics · 1994
Ambiguity is so intrinsically connected to the various levels of linguistic competence that it affects language understanding tasks as well as all related applications (i.e. full‐text retrieval, machine translation). Mainly, in lexical acquisition systems, where increasing attention is devoted to models of language use, the linguistic bootstrapping perspective prevents one from hypothesizing competence models available at the different linguistic levels. This paper describes a generalization of the syntactic model (SSA) described in Basili et al., 1992. The extension is seen to account for some of the morphologic and syntactic ambiguities of the (Italian) language. The notion of elementary syntactic relation (esl) is extended by means of a systematic evaluation of its (syntactic) correctness, i.e. its plausibility. This model will provide a measure of uncertainty as it is dynamically generated by an SSA‐like grammar. Each derived collocate is assigned its plausibility, as a (partial) parse tree is evaluated by a probabilistic grammar. Acquisition based on this source information is shown to produce better lexical information. The syntactic model based on the plausibility measure is also augmented with automatically derived disambiguation rules. The method is demonstrated to be flexible enough to model multiple forms of parsing heuristics. Lexical knowledge derived from different acquisition strategies (frequency vs. plausibility‐driven, blind vs. disambiguated) is evaluated.