Utterance Position-Aware Dialogue Act Recognition

Proceedings of Recent Advances in Natural Language Processing, pages 1567–1574 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_176 1567 Utterance Position-Aware Dialogue Act Recognition Yuki Yano1 Akihiro Tamura2 Takashi Ninomiya3 Hiroaki Obayashi4 1Nara Institute of Science and Technology 2Doshisha University 3Ehime University 4transcosmos inc. yano.yuki.yt0@is.naist.jp aktamura@mail.doshisha.ac.jp ninomiya@cs.ehime-u.ac.jp oobayashi.hiroaki@trans-cosmos.co.jp Abstract This study proposes an utterance position- aware approach for a neural network-based di- alogue act recognition (DAR) model, which in- corporates positional encoding for utterance’s absolute or relative position. The proposed ap- proach 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 en- coding of utterances statistically significantly improves the performance of DAR. 1 Introduction The recognition of dialogue acts (DAs), which rep- resent the intention or function of each utterance in a dialogue, is useful for various dialogue applica- tions, such as dialogue systems and dialogue sum- marization. Recently, neural network- (NN-)based approaches have become dominant in dialogue act recognition (DAR) because NN-based models ob- tained higher performance than other approaches. Most existing NN-based DAR models treat DAR as a sequence labeling problem, where utterances are first encoded with a hierarchical recurrent neural network (RNN), and subsequently, the sequence of DAR labels is identified from the encoded rep- resentations using the Conditional Random Field (CRF) (Kumar et al., 2018; Chen et al., 2018; Li et al., 2019; Raheja and Tetreault, 2019). These CRF-based models can capture the local dependen- cies of DA sequences; however, they cannot model global dependencies due to the first-order Markov assumption. To alleviate this problem, Colombo et al. (2020) proposed a sequence-to-sequence (seq2seq) architecture for DAR, which could learn global dependencies, and achieve the state-of-the- art performance. We focus on this work and employ it as the basis of our proposed model. Existing NN-based DAR models focus on the context of a dialogue and dependencies between DAs but do not focus on the position of an utter- ance. However, we have found that some DAs have tendencies of occurrence positions. For example, the “Open-Question” DA in the Switchboard cor- pus tends to appear at the beginning of a telephonic conversation. Inspired by this observation, this study proposes an utterance position-aware approach for a NN- based DAR model, which explicitly encodes the position of an utterance by positional encoding and generates DA label sequences based on the hidden vectors augmented with the positional encoding of utterance. In particular, we implement two types of positional encodings of utterances: (1) sinusoidal positional encoding used in Transformer neural machine translation (Vaswani et al., 2017), which represents the absolute position of each utterance; (2) length-ratio positional encoding, which is a po- sitional encoding based on the ratio of the position of an utterance to the dialogue length (i.e., the total number of utterances in the dialogue). The sec- ond encoding aims to encode relative positional information to alleviate the variation of dialogue length. The evaluations on the Switchboard corpus (Stol- cke et al., 2000), which is one of the …

Content truncated to 500 words