Semantically Relatable Sets - Building Blocks for Representing Semantics.

Motivated by the fact that automatic analysis of language crucially depends on semantic constituent detection and attachment resolution , we present our work on the problem of generating and linking semantically relatable sets (SRS). These sets are of the form or or , where the entities can be single words or more complex sentence parts (such as embedded clauses). The challenge lies in finding the components of these sets, which involves solving prepositional phrase (PP) and clause attachment problems, and empty pronominal (PRO) determination. Use is made of (i) the parse tree of the sentence, (ii) the subcategorization frames of lexical items, (iii) the lexical properties of the words and (iv) lexical resources like the WordNet and the Oxford Advanced Learners' Dictionary (OALD) . The components within the sets and the sets themselves are linked using the semantic relations of an interlingua for machine translation called the Universal Networking Language (UNL) . The work forms part of a UNL based MT system, where the source language is analysed into semantic graphs and target language is generated from these graphs. The system has been tested on the Penn Treebank , and the results indicate the promise and effectiveness of our approach.