| Presentation | 2007/12/13 The Effect of Competitor and Feature Selection in Discriminative Training of Error Corrective Models Takanobu OBA, Takaaki HORI, Atsushi NAKAMURA, |
|---|---|
| PDF Download Page | PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | We focus on error corrective models for ASR, which select a more accurate word sequence among multiple word sequences produced by a speech recognizer. In such approaches, the corrective model is usually trained discriminatively using hypothesis/reference pairs, typically using word N-gram features to improve discrimination of the reference word sequence from hypothesis word sequences with high recognition scores. However, it is also important for error correction to consider various error patterns. This can be efficiently achieved through the use of hypothesis word sequences with high word error rates (WERs). The various error patterns can be expressed using features that alleviate the data sparseness problem. In this paper, we evaluate the impact of training using competitors with various WERs, as well as the impact of different features, on the corrective model's performance. Our experiments using the Corpus of Spontaneous Japanese show that an accurate and compact model can be generated via training to discriminate the reference from the single worst competing word sequence (with the highest WER),and that models using features that alleviate the data sparseness problem, such as Part-of-Speech N-gram features, achieve robust error correction on evaluation sets. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | error correction / discriminative training / word error rate / competitor / feature selection |
| Paper # | NLC2007-73,SP2007-136 |
| Date of Issue |
| Conference Information | |
| Committee | NLC |
|---|---|
| Conference Date | 2007/12/13(1days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | |
| Vice Chair | |
| Secretary | |
| Assistant | |
| Paper Information | |
| Registration To | Natural Language Understanding and Models of Communication (NLC) |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | The Effect of Competitor and Feature Selection in Discriminative Training of Error Corrective Models |
| Sub Title (in English) | |
| Keyword(1) | error correction |
| Keyword(2) | discriminative training |
| Keyword(3) | word error rate |
| Keyword(4) | competitor |
| Keyword(5) | feature selection |
| 1st Author's Name | Takanobu OBA |
| 1st Author's Affiliation | NTT Communication Science Laboratories, NTT Corporation() |
| 2nd Author's Name | Takaaki HORI |
| 2nd Author's Affiliation | NTT Communication Science Laboratories, NTT Corporation |
| 3rd Author's Name | Atsushi NAKAMURA |
| 3rd Author's Affiliation | NTT Communication Science Laboratories, NTT Corporation |
| Date | 2007/12/13 |
| Paper # | NLC2007-73,SP2007-136 |
| Volume (vol) | vol.107 |
| Number (no) | 405 |
| Page | pp.pp.- |
| #Pages | 6 |
| Date of Issue | |