The growing demand for mental health support has highlighted the importance of conversational agents as human supporters worldwide and in China. These could increase availability reduce relative costs support. provided can be divided into two main types: cognitive emotional. Existing work on this topic mainly focuses constructing that adopt Cognitive Behavioral Therapy (CBT) principles. Such operate based pre-defined templates exercises to provide However, research emotional using such is...

Endowing dialogue systems with personas is essential to deliver more human-like conversations. However, this problem still far from well explored due the difficulties of both embodying personalities in natural languages and persona sparsity issue observed most corpora. This paper proposes a pre-training based personalized model that can generate coherent responses using persona-sparse data. In method, pre-trained language used initialize an encoder decoder, personal attribute embeddings are...

Natural Language Understanding (NLU) is a vital component of dialogue systems, and its ability to detect Out-of-Domain (OOD) inputs critical in practical applications, since the acceptance OOD input that unsupported by current system may lead catastrophic failure. However, most existing detection methods rely heavily on manually labeled samples cannot take full advantage unlabeled data. This limits feasibility these models applications. In this paper, we propose novel model generate...

Endowing a dialogue system with particular personality traits is essential to deliver more human-like conversations. However, due the challenge of embodying via language expression and lack large-scale persona-labeled data, this research problem still far from well-studied. In paper, we investigate incorporating explicit in generation personalized dialogues. To end, firstly, construct PersonalDialog, multi-turn dataset containing various large number speakers. The consists 20.83M sessions...

Pre-trained models have proved to be powerful in enhancing task-oriented dialog systems. However, current pre-training methods mainly focus on understanding and generation tasks while neglecting the exploitation of policy. In this paper, we propose GALAXY, a novel pre-trained model that explicitly learns policy from limited labeled dialogs large-scale unlabeled corpora via semi-supervised learning. Specifically, introduce act prediction task for optimization during employ consistency...

Pre-trained models have proved to be powerful in enhancing task-oriented dialog systems. However, current pre-training methods mainly focus on understanding and generation tasks while neglecting the exploitation of policy. In this paper, we propose GALAXY, a novel pre-trained model that explicitly learns policy from limited labeled dialogs large-scale unlabeled corpora via semi-supervised learning. Specifically, introduce act prediction task for optimization during employ consistency...

Although pre-trained language models have remarkably enhanced the generation ability of dialogue systems, open-domain Chinese systems are still limited by data and model size compared with English ones. In this paper, we propose EVA, a system that contains largest 2.8B parameters. To build model, collect dataset named WDC-Dialogue from various public social media. This 1.4B context-response pairs is used as pre-training corpus EVA. Extensive experiments on automatic human evaluation show EVA...

Yingxiu Zhao, Zhiliang Tian, Huaxiu Yao, Yinhe Zheng, Dongkyu Lee, Yiping Song, Jian Sun, Nevin Zhang. Proceedings of the 60th Annual Meeting Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Recent advances in open-domain dialogue systems rely on the success of neural models that are trained large-scale data. However, collecting data is usually time-consuming and labor-intensive. To address this dilemma, we propose a novel augmentation method for training by utilizing unpaired Specifically, data-level distillation process first proposed to construct augmented dialogues where both post response retrieved from A ranking module employed filter out low-quality dialogues. Further,...

Yida Wang, Yinhe Zheng, Yong Jiang, Minlie Huang. Proceedings of the 59th Annual Meeting Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Empathetic dialogue is a human-like behavior that requires the perception of both affective factors (e.g., emotion status) and cognitive cause emotion). Besides concerning status in early work, latest approaches study causes empathetic dialogue. These focus on understanding duplicating context to show empathy for speaker. However, instead only repeating contextual causes, real empathic response often demonstrate logical emotion-centered transition from those responses. In this we propose an...

In traditional block theory, the removability and stability of rock blocks are analyzed independently; that is, a removable is in detail, nonremovable regarded as stable. However, practical situations, may pose more danger than blocks. This paper presents unified method for analyzing this method, cracking bridges considered not assumed to be First, possible identified by extending finite-sized fractures comparing boundary surfaces resulting with those original Then, sliding direction...

Generating stylized responses is essential to build intelligent and engaging dialogue systems. However, this task far from well-explored due the difficulties of rendering a particular style in coherent responses, especially when target embedded only unpaired texts that cannot be directly used train model. This paper proposes generation method can capture stylistic features texts. Specifically, our produce are both given context conform style. In study, an inverse model first introduced...

Out-of-distribution (OOD) detection is essential for the reliable and safe deployment of machine learning systems in real world. Great progress has been made over past years. This paper presents first review recent advances OOD with a particular focus on natural language processing approaches. First, we provide formal definition discuss several related fields. We then categorize algorithms into three classes according to data they used: (1) available, (2) unavailable + in-distribution (ID)...

Grounded dialogue models generate responses that are grounded on certain concepts. Limited by the distribution of data, trained such data face transferability challenges in terms and type To address challenges, we propose minimal editing framework, which minimally edits existing to be given concept. Focusing personas, Minimal Editor (GME), learns edit disentangling recombining persona-related persona-agnostic parts response. evaluate persona-grounded editing, present PersonaMi-nEdit dataset,...

Lifelong learning (LL) is vital for advanced task-oriented dialogue (ToD) systems. To address the catastrophic forgetting issue of LL, generative replay methods are widely employed to consolidate past knowledge with generated pseudo samples. However, most existing use only a single task-specific token control their models. This scheme usually not strong enough constrain model due insufficient information involved. In this paper, we propose novel method, prompt conditioned VAE lifelong...

Out-of-Domain (OOD) intent detection is important for practical dialog systems. To alleviate the issue of lacking OOD training samples, some works propose synthesizing pseudo samples and directly assigning one-hot labels to these samples. However, introduce noises process because "hard" may coincide with In-Domain (IND) intents. In this paper, we an adaptive soft labeling (ASoul) method that can estimate when detectors. Semantic connections between IND intents are captured using embedding...

Incorporating multi-modal contexts in conversation is important for developing more engaging dialogue systems. In this work, we explore direction by introducing MMChat: a large-scale Chinese corpus (32.4M raw dialogues and 120.84K filtered dialogues). Unlike previous corpora that are crowd-sourced or collected from fictitious movies, MMChat contains image-grounded real conversations on social media, which the sparsity issue observed. Specifically, image-initiated common communications may...

Unsupervised domain adaptation (UDA) with pre-trained language models (PrLM) has achieved promising results since these embed generic knowledge learned from various domains. However, fine-tuning all the parameters of PrLM on a small domain-specific corpus distort knowledge, and it is also expensive to deployment whole fine-tuned for each domain. This paper explores an adapter-based approach unsupervised adaptation. Specifically, several trainable adapter modules are inserted in PrLM,...

This paper proposes a dual-supervised uncertainty inference (DS-UI) framework for improving Bayesian estimation-based UI in DNN-based image recognition. In the DS-UI, we combine classifier of DNN, i.e., last fully-connected (FC) layer, with mixture Gaussian models (MoGMM) to obtain an MoGMM-FC layer. Unlike existing methods DNNs, which only calculate means or modes DNN outputs' distributions, proposed layer acts as probabilistic interpreter features that are inputs directly probabilities...