Current studies on the generation of personalized dialogue primarily contribute to an agent presenting a consistent personality and driving a more informative response. However, we found that the responses generated from most previous models were self-centered, with little consideration for the user in the dialogue. Moreover, we consider human-like conversations to be essentially based on inferring information about the persona of the other party. Therefore, we propose a novel personalized dialogue generator that detects implicit user personas. Because it is difficult to collect a large amount of detailed personal facts for each user, we attempted to model the potential persona of a user and its representation from the dialogue history with no external knowledge. The perception and fader variables were conceived using conditional variational inference. The two latent variables simulate the process of people becoming aware of each other’s personas and producing a corresponding expression in conversations. Subsequently, posterior-discriminated regularization was performed to enhance the training procedure. Finally, a selector was designed to help our model provide long-sighted responses. Comprehensive experiments demonstrate that compared to the state-of-the-art methods, our approach is more concerned with the user’s persona and achieves a notable boost across both automatic metrics and human evaluations.
Text infilling aims to restore incomplete texts by filling in blanks and has attracted increasing attention recently because of its wide application in ancient text restoration, conversation generation, and text rewriting. However, attribute-aware text infilling is yet to be explored, and existing methods seldom focus on the infilling length of each blank or the number and location of the blanks. In this study, we propose a plug-and-play Attribute-aware Text Infilling method using a Pre-trained language model (A-TIP) that contains a text-infilling component and a plug-and-play discriminator. Specifically, we first designed a unified text-infilling component with modified attention mechanisms and intra- and inter-blank positional encoding to better perceive the number of blanks and the infilling length for each blank. We then propose a plug-and-play discriminator to guide generation and improve attribute relevance without decreasing text fluency. Finally, automatic and human evaluations on three open-source datasets indicate that, compared to all the baselines, A-TIP achieves state-of-the-art performance. An additional ablation study demonstrated the robustness of A-TIP.
We introduce the task of implicit offensive-text detection (OTD) in dialogues, where a statement may have either an offensive or nonoffensive interpretation depending on the listener and context. We argue that reasoning is crucial for understanding this broader class of offensive utterances, and release SLIGHT, a test dataset to support research on this topic. Experiments using the data show that state-of-the-art methods for offense detection perform poorly when tasked with detecting implicitly offensive statements, achieving only ∼11% accuracy. In contrast to the existing OTD datasets, SLIGHT features human-annotated chains of reasoning that describe the mental process through which an offensive interpretation can be reached from an ambiguous statement. We explore the potential of a multihop reasoning approach, by utilizing the existing entailment models to evaluate the probabilities of these chains. Our results demonstrate that reasoning through chains can yield performances better than that of a baseline entailment setting without chains. Furthermore, the analysis of the chains provides insights into the human interpretation process and emphasizes the importance of incorporating additional commonsense knowledge.