Improved speech understanding using dialogue expectation in sentence parsing
Sherif Mahdy Abdou, Michael S. Scordilis · 2000
In dialogue systems, speech recognition errors force the user to repeat information resulting in more turns, lower dialogue efficiency and maybe complete failure. Higher level information such as history, expectation, discourse knowledge and pragmatics can improve performance but are hard to quantify and effectively include in the recognition process. In this paper, dialogue expectation is used to improve recognition. By permitting the dialogue manager to guide the interaction it is possible to track the dialogue state and thus estimate the expected semantic content of the user’s response. The parser is allowed to process a large number of sentences provided by the decoder. Expectation is used as an effective criterion for selecting among competing hypotheses. This approach was tested with a simple flight reservation task and results show improvement in concept recognition without adding significant computation.