Utility of the representation

The approach to lexical semantics within HPSG as outlined above has several positive properties.

1. Semantic relations are constrained according to their definition in the subsumption hierarchy. This definition specifies how many (semantic) arguments the relation has and associates each argument with a particular semantic role.
2. Semantic relation types are organised via an inheritance hierarchy which allows similar types to be grouped together. Generalisations over these groups can therefore be directly represented in the hierarchy.
3. The semantic roles filled by the syntactic arguments of a word can be specified directly in the lexical entry of the word, via structure-sharing.

However, there are also some shortcomings to the system as presented in Pollard and Sag (1994), mainly due to the fact that a complete description of lexical semantic representation was by no means a goal of that work. The shortcomings identified here relate directly to the problem of efficiently capturing sense extensions which is the main focus of this thesis.

1. Alternations in the surface order of syntactic arguments would have to be captured by distinct lexical entries. These lexical entries would differ only in the mapping between the syntactic arguments and the semantic arguments, thereby missing subtle semantic distinctions between the alternate forms.

   The alternate forms could be generated via lexical rules, but this would either require one form to be the base form (see arguments against this in Markantonatou and Sadler marks_sadler:95), or would require one lexical rule to generate each of the alternate forms from a verbal stem in the lexicon. Whichever approach is chosen, these lexical rules would not be constrained enough given the existing grain of semantic representation in HPSG to block the application of the rules to non-alternating verbs.

2. Each sense of a word must correspond to a distinct semantic relation type in the sort hierarchy, and must be defined by a lexical entry specific to that sense. This fails to adequately capture relationships among the senses. Polysemous words, for which the different senses of the verb share a certain amount of meaning, are treated equivalently to truly ambiguous words, for which one word corresponds to multiple entirely distinct meanings (e.g. bank as in edge of a river and financial institution). This problem could be addressed by introducing polymorphic types (along the lines of e.g. Sanfilippo 1995).
3. No allowance is made for the specification of selectional restrictions, or for reference to the kinds of entities which instantiate the semantic arguments of the relations. This is a problem for, for example, a proper treatment of logical metonymy, to be addressed in Chapter 5, which requires information about the referent of a complement noun in order to appropriately specify the psoa expressed by a sentence.

These problems could be alleviated through the incorporation of a certain amount of semantic decomposition, along the lines of what Jackendoff (1983, 1990) has proposed for verb semantics and what Pustejovsky (1991, pustejovsky:95a) has proposed for nominal semantics.