Utterance Position-Aware Dialogue Act Recognition

Abstract

This study proposes an utterance positionaware approach for a neural network-based dialogue act recognition (DAR) model, which incorporates positional encoding for utterance's absolute or relative position. The proposed approach is inspired by the observation that some dialogue acts have tendencies of occurrence positions. The evaluations on the Switchboard corpus show that the proposed positional encoding of utterances statistically significantly improves the performance of DAR.

Surface and Contextual Linguistic Cues in Dialog Act Classification: A Cognitive Science View

“…Finally, there is a marginal influence of utterance position and speaker cues in both analyses. This corresponds to findings in the dialog act classification literature, where marginal increases of up to 2% were found when speaker cues (Zhao & Kawahara, 2019), utterance position cues (Yano et al, 2021), and a combination of speaker, utterance position, and utterance length cues (Li & Wu, 2016) were added to a baseline of surface linguistic cues. In fact, as we observed in the literature overview, many deep learning approaches achieve a high performance without considering the structure of turns, nor who uttered which utterance.…”

“…For the same reasons, it might be surprising that turn and discourse organization cues have seldomly been used in dialog act classification studies. In fact, information on the location of the utterance in the turn can indeed be beneficial for dialog act classification (Yano et al, 2021). Finally, specific to the Map Task corpus, Di Eugenio et al (2010) have annotated dialogs with a label denoting the subdialog they are part of.…”

“…Context is either modeled through a condensed vector of contextual cues (e.g., Chen et al, 2018;Colombo et al, 2020;Kumar et al, 2018;Raheja & Tetreault, 2019) or directly in the neural network structure with a node for each utterance (e.g., Cerisara et al, 2018;Kalchbrenner & Blunsom, 2013;Ortega & Vu, 2017;Ribeiro et al, 2019b). Some studies that used deep learning for dialog act classification did not include contextual cues (e.g., Duran & Battle, 2018;Duran et al, 2023;Khanpour et al, 2016;Ribeiro et al, 2019a), but most encoded context through using a condensed surface encoding of the previous utterances (e.g., Chen et al, 2018;Kumar et al, 2018;Yano et al, 2021;Zhao & Kawahara, 2019). Most deep learning models have regarded dialog act classification as classifying a sequence of dialog acts, without paying attention to the speaker or the structure of utterances into turns.…”

“…Most deep learning models have regarded dialog act classification as classifying a sequence of dialog acts, without paying attention to the speaker or the structure of utterances into turns. Some more recent deep learning approaches have seen success with incorporating some form of incorporating speaker cues (e.g., Li & Wu, 2016;Ribeiro et al, 2019b;Zhao & Kawahara, 2019) and incorporating the turn structure (Li & Wu, 2016;Yano et al, 2021). Finally, many of the more recent deep learning models included some form of an attention mechanism (e.g., Li et al, 2019;Raheja & Tetreault, 2019;Tran et al, 2017a;Wan et al, 2018).…”

“…In both these studies, the models were evaluated on the Map Task corpus where the instruction giver and instruction follower might use words differently. In (other) machine learning and deep learning approaches, speaker cues have typically been encoded using a feature that tells the model who the speaker is (Bothe, Weber et al., 2018 ; Cerisara et al., 2018; Di Eugenio et al., 2010) or whether the previous utterance is uttered by the same or different speaker from the current utterance (e.g., Lager & Zinovjeva, 1999; Liu et al., 2017; Yano et al., 2021; Zhao & Kawahara, 2019). Interestingly, many deep learning models, despite encoding contextual cues such as the structure of turns directly into the model, do not take into account who uttered which utterance, and still achieve a good performance.…”