Ensemble application of symbolic and subsymbolic AI for open-domain sentiment-aware dialogue systems

Abstract

With the development of AI techniques, dialogue systems have attracted the attention of the research community and become a main research focus in natural language processing(NLP) due to their promising potential and application values. Dialogue and dialogue systems refer to the human-to-human and human-robot conversations in this thesis. Dialogues that extend beyond a narrow range of topics are categorized as open-domain dialogues. People usually express their opinions or sentiments in dialogue or dialogue systems. Consequently, emotion recognition in conversations(ERC) and sentiment interaction identification between speakers are crucial for understanding sentiment-aware dialogue systems. However, the complex context, multiple participants and informal text make it hard to accurately identify the emotion of each utterance in conversations. To this end, we proposed two solutions for addressing the above issues in the ERC task and a new paradigm for an in-depth understanding of the sentiment interactions in conversations.

In contrast to the traditional sentiment analysis task, the primary challenge in the ERC task lies in handling contextual information as context information sometimes enhances, weakens or reverses the raw emotion of an utterance. Hence, we proposed a generalized neural tensor block for the ERC task to better integrate the context information into the representation of the current utterance. With context-aware utterance features, we successfully transformed the ERC task into a conventional sentiment analysis task. Then a two-channel feature extractor followed by a softmax layer is designed to classify the context-aware features. An empirical study on three benchmark datasets demonstrated the effectiveness and superiority of the proposed framework. Aside from its good performance, the framework also demonstrates efficiency in terms of both parameters and computation.

Furthermore, to investigate the functionality and interpretability of the proposed neural tensor block, we presented a mathematical explanation of the neural tensor network, revealing the inner link between the feedforward neural network and neural tensor network and providing factorization methods to make neural tensor network parameter-efficient. Extensive experiments on three typical NLP tasks including the ERC proved the correctness of Taylor's theorem-based explanation.

Thanks to the rapid development of the pre-trained language model (PLM), we observed significant performance gains in the ERC task recently. However, these neuro-based sub-symbolic systems still have some limitations, e.g., lack of explainability, poor reasoning ability, and untrustworthy results. Besides, we observed that some utterances seldom contain sentiment information and dependencies widely exist between utterances. To address the challenges in the ERC task and limitations of PLM, we proposed a neural-symbolic framework named SKIER fusing different symbolic knowledge and PLM representation to achieve better performance and explainability. An empirical study on benchmarks demonstrated the effectiveness of the neural-symbolic framework as well as the functionality of the symbolic module. While the ERC task helps in comprehending the emotional state of the speakers, the cause behind a certain emotion remains unclear. The emotion and its corresponding cause utterances reflect the emotional interaction between speakers, i.e., how does one speaker influence the emotion of another speaker? To this end, we introduced a new task named emotion-cause pair extraction in conversations (ECPEC) to extract all possible emotion-cause utterance pairs in conversations. To the best of our knowledge, there is no existing dataset for the ECPEC task, hence we built a high-quality task-specific dataset named ConvECPE. The ECPEC task is more challenging than the ERC as dozens of emotion-cause pairs need to be filtered from thousands of candidate emotion-cause pairs. Hence, we provided a two-stage framework comprising stage one: a multitask learning extractor to jointly extract emotion and cause utterances, and stage two: a filter with emotion-cause-chunk pair as an auxiliary task to filter candidate emotion-cause pairs. Experiments on the ConvECPE dataset demonstrated the feasibility and effectiveness of ECPEC.

In conclusion, this thesis addressed some challenging problems in recognizing the emotions in conversations and introduced a new paradigm for promoting profound comprehension of sentiment-aware dialogue systems. Besides, the influence of large language models like ChatGPT on the research of sentiment-aware dialogue systems is also examined in this thesis.