Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
This paper proposes a topic indentifier which uses a multi-layer perceptron (MLP) with a keyword-spotting preprocessor. The keyword-spotting preprocessor is designed to locate the keywords that appear in the input sentences and assigns each of them to one class among a pre-determined set of semantic classes. A three layer perceptron is then trained to identify the input sentence's topic, the network's input units represent semantic classes and its output units correspond to the topics.
The experiments on AITS0, the pilot corpus from the Air Travel Information System (ATIS) corpora, show 91.01% classifying accuracy for the test set when transcripts are used as input, and 88.76% accuracy when using the output from CRIM's spontaneous speech recognition system[l].
Bibliographic reference. Cheng, Ying / Fortier, Paul / Normandin, Yves (1992): "Topic identification using a neural network with a keyword-spotting preprocessor", In ICSLP-1992, 1609-1612.