Generating persona consistent dialogue response is important for developing an intelligent conversational agent. Recent works typically fine-tune large-scale pre-trained models on this task by concatenating persona texts and dialogue history as a single input sequence to generate the target response. While simple and effective, our analysis shows that this popular practice is seriously affected by order sensitivity where different input orders of persona sentences significantly impact the quality and consistency of generated response, resulting in severe performance fluctuations (i.e., 29.4% on GPT2 and 83.2% on BART). To mitigate the order sensitivity problem, we propose a model-agnostic framework, ORder Insensitive Generation (ORIG), which enables dialogue models to learn robust representation under different persona orders and improve the consistency of response generation. Experiments on the Persona-Chat dataset justify the effectiveness and superiority of our method with two dominant pre-trained models (GPT2 and BART).

Humor plays an important role in our daily life, as it is an essential and fascinating element in the communication between persons. Therefore, how to recognize punchlines from the dialogue, i.e. conversational humor recognition, has attracted much interest of computational linguistics communities. However, most existing work attempted to understand the conversational humor by analyzing the contextual information of the dialogue, but neglected the character of the interlocutor, such as age, gender, occupation, and so on. For instance, the same utterance could bring out humorous from a serious person, but may be a plain expression from a naive person. To this end, this paper proposes a Character Fusion Conversational Humor Recognition model (CFCHR) to explore character information to recognize conversational humor. CFCHR utilizes a multi-task learning framework that unifies two highly pertinent tasks, i.e., character extraction and punchline identification. Based on deep neural networks, we trained both tasks jointly by sharing weight to extract the common and task-invariant features while each task could still learn its task-specific features. Experiments were conducted on Chinese sitcoms corpus, which consisted of 12,677 utterances from 22 characters. The experimental results demonstrated that CFCHR could achieve 33.08% improvements in terms of F1-score over some strong baselines, and proved the effectiveness of the character information to identify the punchlines.