Research On Recognition Of Underrepresented Named Entities In Speech Recognition

In recent years,human-computer speech interaction(HCSI)has been more and more concerned and loved by people.Among them,automatic speech recognition(ASR)plays an important role in HCSI system.In ASR service,it is often necessary to recognize named entity(NE),such as: person name,organization,country,etc.However,these NEs are easily misrecognized by the system.Especially for some underrepresented named entities(UR-NE),here UR-NE means that the frequency of NE in training data is less than 10 or does not appear in training data,the performance of ASR system will decrease sharply.Meanwhile,the wrong recognition of these NEs can lead to the failure of some downstream applications,and the efficiency of HCSI will also be reduced.In ASR,it is a challenging task to recognize the UR-NEs due to such NEs have insufficient instances and poor contextual coverage in the training data to learn reliable estimates and representations.In order to improve the recognition accuracy of UR-NEs in speech recognition,this paper proposes a series of methods based on the framework of HMM speech recognition system for research and analysis.In view of the high cost of making phoneme pronunciation lexicon in the speech recognition system and the problems that grapheme-to-phoneme(G2P)will introduce some errors when processing out-of-vocabulary(OOV).Especially for some UR-NEs,it is difficult for G2P to learn reliable pronunciation sequences because of their irregular pronunciation.Therefore,this paper employs grapheme pronunciation lexicon to replace the phoneme pronunciation lexicon that commonly used in speech recognition system to connect acoustic model(AM)and language mode(LM).On this basis,this paper proposes a series of methods to enrich the representations of UR-NEs to improve the recognition performance of speech recognition system.Specifically,this paper first ensures those UR-NEs appear in the word lattice generated by first-pass decoding.To this end,this paper employs class-based LM philosophy,making exemplar utterances for those UR-NEs according to their classes.By using these exemplar utterances to update an N-gram LM that boosts the UR-NE occurrence in the word lattice.Then,in the second-pass,this paper proposes two lattice rescoring methods to further improve the recognition performance of the system.Among them,the first rescoring is to enrich the representations of UR-NEs in a pretrained recurrent neural network LM(RNNLM)by borrowing the embedding representations of richrepresented NEs(RR-NE),yielding the lattice that statistically favor the UR-NEs.The second rescoring directly boosts the likelihood scores of the utterances containing UR-NEs in the lattice and gain further performance improvement.Finally,the combination of these methods significantly improves the recognition performance of UR-NEs in speech recognition.