Hybrid Syntactic-Semantic Reranking for Parsing Results of ECAs Interactions using CRFs

Reranking modules of conventional parsers make use of either probabilistic weights linked to the production rules or just hand crafted rules to choose the best possible parse. Other proposals make use of the topology of the parse trees and lexical features to reorder the parsing results. In this work, a new reranking approach is presented. There are two main novelties introduced in this paper: firstly, a new discriminative reranking method of parsing results has been applied using Conditional Random Fields (CRFs) for sequence tagging. Secondly, a mixture of syntactic and semantic features, speci cally designed for Embodied Conversational Agents (ECAs) interactions, has been used. This approach has been trained with a Corpus of over 4,000 dialogues, obtained from real interactions of real users with an online ECA. Results show that this approach provides a significant improvement over the parsing results of out-of-domain sentences; that is, sentences for which there is no optimal parse among the candidates given by the baseline parse.