Recent development in NLP shows a strong trend towards refining pre-trained models with a domain-specific dataset. This is especially the case for response generation where emotion plays an important role. However, existing empathetic datasets remain small, delaying research efforts in this area, for example, the development of emotion-aware chatbots. One main technical challenge has been the cost of manually annotating dialogues with the right emotion labels. In this paper, we describe a large-scale silver dataset consisting of 1M dialogues annotated with 32 fine-grained emotions, eight empathetic response intents, and the Neutral category. To achieve this goal, we have developed a novel data curation pipeline starting with a small seed of manually annotated data and eventually scaling it to a satisfactory size. We compare its quality against a state-of-the-art gold dataset using both offline experiments and visual validation methods. The resultant procedure can be used to create similar datasets in the same domain as well as in other domains.

Empathetic dialog generation aims at generating coherent responses following previous dialog turns and, more importantly, showing a sense of caring and a desire to help. Existing models either rely on pre-defined emotion labels to guide the response generation, or use deterministic rules to decide the emotion of the response. With the advent of advanced language models, it is possible to learn subtle interactions directly from the dataset, providing that the emotion categories offer sufficient nuances and other non-emotional but emotional regulating intents are included. In this paper, we describe how to incorporate a taxonomy of 32 emotion categories and 8 additional emotion regulating intents to succeed the task of empathetic response generation. To facilitate the training, we also curated a large-scale emotional dialog dataset from movie subtitles. Through a carefully designed crowdsourcing experiment, we evaluated and demonstrated how our model produces more empathetic dialogs compared with its baselines.

In this paper, we describe our system submitted to SemEval 2021 Task 7: HaHackathon: Detecting and Rating Humor and Offense. The task aims at predicting whether the given text is humorous, the average humor rating given by the annotators, and whether the humor rating is controversial. In addition, the task also involves predicting how offensive the text is. Our approach adopts the DeBERTa architecture with disentangled attention mechanism, where the attention scores between words are calculated based on their content vectors and relative position vectors. We also took advantage of the pre-trained language models and fine-tuned the DeBERTa model on all the four subtasks. We experimented with several BERT-like structures and found that the large DeBERTa model generally performs better. During the evaluation phase, our system achieved an F-score of 0.9480 on subtask 1a, an RMSE of 0.5510 on subtask 1b, an F-score of 0.4764 on subtask 1c, and an RMSE of 0.4230 on subtask 2a (rank 3 on the leaderboard).

Humor recognition has been widely studied as a text classification problem using data-driven approaches. However, most existing work does not examine the actual joke mechanism to understand humor. We break down any joke into two distinct components: the set-up and the punchline, and further explore the special relationship between them. Inspired by the incongruity theory of humor, we model the set-up as the part developing semantic uncertainty, and the punchline disrupting audience expectations. With increasingly powerful language models, we were able to feed the set-up along with the punchline into the GPT-2 language model, and calculate the uncertainty and surprisal values of the jokes. By conducting experiments on the SemEval 2021 Task 7 dataset, we found that these two features have better capabilities of telling jokes from non-jokes, compared with existing baselines.