24.5.3 Detecting Correction Acts

In addition to general-purpose dialogue act interpretation, we may want to build special-purpose detectors for particular acts. Let's consider one such detector, for the recognition of user correction of system errors. If a dialogue system misrecognizes an utterance (usually as a result of ASR errors) the user will generally correct the error by repeating themselves, or rephrasing the utterance. Dialogue systems need to recognize that users are doing a correction, and then figure out what the user is trying to correct, perhaps by interacting with the user further.

Unfortunately, corrections are actually harder to recognize than normal sentences. Swerts et al. (2000) found that corrections in the TOOT dialogue system were misrecognized about twice as often (in terms of WER) as non-corrections. One reason for this is that speakers use a very different prosodic style called hyperarticulation for corrections. In hyperarticulated speech, some part of the utterance has exaggerated energy, duration, or F0 contours, such as I said BAL-TI-MORE, not Boston (Wade et al., 1992; Oviatt et al., 1998; Levow, 1998; Hirschberg et al., 2001).

原书第 969 页

Figure 24.19 Decision tree for classifying DECLARATIVE QUESTIONS (QD), STATEMENT (S), YES-NO QUESTIONS (QY), and WH-QUESTIONS (QW), after Shriberg et al. (1998). Each node in the tree shows four probabilities, one for each of the four dialogue acts in the order QD, S, QY, QW; the most likely of the four is shown as the label for the node. Note that questions are shorter than statements (cont_speech_frames), that QY's rise at the end (end_grad), and that QD's rise throughout the utterance (utt_grad).

Speakers use various strategies to make corrections; Swerts et al. (2000) categorize these as repetition (REP), paraphrase (PAR), adding content (ADD), and omitting content (OMIT), as shown in Fig. 24.20. They found that most corrections were repetitions and omissions rather than paraphrases.

Various machine learning classifiers have been applied to the problem of detecting correction acts. Useful features include lexical information (words like “no”, “correction”, “I don't”, swear words), prosodic and hyperarticulation features (increases in F0 range, pause duration, and word duration, generally normalized by the values for previous sentences), features indicating utterance length, ASR features (confidence, language model probability), and various dialogue features (Levow, 1998; Hirschberg et al., 2001; Bulyko et al., 2005).

In addition to correction detection, a conversational agent also needs appropriate control or update rules in the dialogue manager (Bulyko et al., 2005).