The work presented in this paper regards the construction of the formal representation of the content of the input sentence. It is part of a speech-to-speech translation system, where this language independent format is used as an intermediate representation between source and target languages. The general application domain is travel planning, but it is divided in more specific domains. such as hotel and transport reservation. Starting from the word string output by the recognizer, the construction of the content representation requires the recognition of: 1) the speech act and concept list, and 2) a list of arguments in the form of argument name/value pair. For the former, semantic classification trees [5] are used and domain portability depends on the availability of labeled data for the new domain. For argument extraction a pipeline of filters is used, where each filter is based on a semantic grammar, which in most cases is regular. Then, these filters can be used all together for the most general analysis, or separately when only one or a few domains are considered. Argument extraction is described and evaluated in terms of the changes in precision and recall with respect to the number of filters and of rules
An Inter-Domain Portable Approach to Interchange Format Construction
1999-01-01
Abstract
The work presented in this paper regards the construction of the formal representation of the content of the input sentence. It is part of a speech-to-speech translation system, where this language independent format is used as an intermediate representation between source and target languages. The general application domain is travel planning, but it is divided in more specific domains. such as hotel and transport reservation. Starting from the word string output by the recognizer, the construction of the content representation requires the recognition of: 1) the speech act and concept list, and 2) a list of arguments in the form of argument name/value pair. For the former, semantic classification trees [5] are used and domain portability depends on the availability of labeled data for the new domain. For argument extraction a pipeline of filters is used, where each filter is based on a semantic grammar, which in most cases is regular. Then, these filters can be used all together for the most general analysis, or separately when only one or a few domains are considered. Argument extraction is described and evaluated in terms of the changes in precision and recall with respect to the number of filters and of rulesI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
