2.2. Confidence Baseline
A common approach to improve NER performance on speech is to incorporate a confidence estimate for predicting decoding errors (such as those caused by OOVs) [3, 4] . Therefore, we compare our app roach with both a baseline NER featu resetde- scribed in Section 2.1, and with an additional confidence esti- mat e baseline (Figure 1.) The confidence baseline uses the features of Sudoh et al. [4] to create a CRF error predictor: the decoded word, POS tag, and posterior probability, as well as these features from a ±2 word window. This system shows superior error detection performance to only using the word posterior probability. The training data was obtained from a standard word-based LVCSR system whose errors are known by aligning with the reference transcription. The probability of error, provided by the CRF error predictor, was quantized into 10 bins generating binary features (errordet).
đang được dịch, vui lòng đợi..