Linking Semantic and Knowledge Representations in a Multi-Domain Dialogue System

M. O. Dzikovska, J. F. Allen, M. D. Swift
2007 Journal of Logic and Computation  
We describe a two-layer architecture for supporting semantic interpretation and domain reasoning in dialogue systems. Building systems that support both semantic interpretation and domain reasoning in a transparent and well-integrated manner is an unresolved problem because of the diverging requirements of the semantic representations used in contextual interpretation versus the knowledge representations used in domain reasoning. We propose an architecture that provides both portability and
more » ... ciency in natural language interpretation by maintaining separate semantic and domain knowledge representations, and integrating them via an ontology mapping procedure. The ontology mapping is used to obtain representations of utterances in a form most suitable for domain reasoners, and to automatically specialize the lexicon. The use of a linguistically motivated parser for producing semantic representations for complex natural language sentences facilitates building portable semantic interpretation components as well as connections with domain reasoners. Two evaluations demonstrate the effectiveness of our approach: we show that a small number of mapping rules is sufficient for customizing the generic semantic representation to a new domain, and that our automatic lexicon specialization technique improves parser speed and accuracy. the user's actions, which requires relating natural language to non-linguistic knowledge sources such as databases, reasoners and planners. Effective language processing in dialogue systems involves the production of representations for natural language that can be used both in semantic interpretation and in efficient domain reasoning. Sophisticated semantic representations such as UDRS [24] or MRS [15] are supported by deep grammars such as LINGO ERG [14] or XLE [37] . However, these semantic representations are limited to features that can be reliably extracted from syntax, which presents difficulties for their application in domain reasoning tasks (see Section 2.2 for discussion). As a result, practical systems using these representations have been implemented only in small domains requiring fairly simple domain knowledge, such as appointment scheduling [30] . Dialogue systems that use more complex domain reasoners tend to rely on either on domain-specific grammars, or on wider-coverage syntactic grammars with domain-specific lexicons [42, 33, 17, 28] . However, the language components in such dialogue systems are difficult to port to new domains because of their domain-specific representations. They are also not easily extended to handle complex phenomena in semantic interpretation because the knowledge representation languages used for efficient reasoning are not well-equipped to represent the features needed for contextual interpretation, for example, underspecification or representations for context-dependent entities such as pronouns. This division reflects a general distinction between linguistic and pragmatic knowledge and domain knowledge. Linguistic and pragmatic knowledge is generally domainindependent. Many algorithms for reference resolution [10], intention recognition [8], dialogue management [32] and fragment interpretation [44] are formulated using generic notions such as plans and resources. They may need to call on domain reasoners to verify that something is a resource or a plan in the given domain, but the core of the algorithm uses general knowledge about dialogue that does not change between applications. Domain reasoners, however, are built to encode knowledge and support efficient queries and inferences within a given domain. We address the problem of how to build parsers and grammars that support the diverging representation requirements imposed by those tasks. We propose a parsing and interpretation system architecture that combines the benefits of a wide-coverage deep parser for semantic interpretation with customization techniques to support efficient domain reasoning. The parser builds representations that support semantic interpretation in multiple domains using a lexicon linked to a domaingeneral language ontology (constructed with the aid of several task-oriented dialogue corpora). The domain-general semantic representations are linked with knowledge representations in specific domains via an ontology mapping technique. Maintaining a domain-general grammar and lexicon facilitates building interpretation components that are easy to port across domains, while our ontology mappings provide customized connections to the domain knowledge representation. Moreover, we develop an automatic method of specializing the lexicon based on the ontology mappings that significantly improves parsing speed and accuracy for in-domain sentences, while maintaining portability and reusability in the grammar and lexicon. We demonstrate the effectiveness of this approach by evaluating the performance of the same wide-coverage grammar and lexicon in two different domains. We show that the parser performs similarly in both domains, and that a small number of mapping rules Parser
doi:10.1093/logcom/exm067 fatcat:vggzrjz23rbljdklumbibfr4ma