Source-linked AI summary
CEM: Commonsense-aware Empathetic Response Generation
Sahand Sabour, Chujie Zheng, Minlie Huang
TL;DR
Existing empathetic dialogue methods emphasize detecting users' emotions, although empathy also requires cognitive understanding of their situations. CEM uses commonsense knowledge to infer additional situational information and reports more informative and empathetic responses in automatic and manual evaluations.
Problem
Previous empathetic dialogue methods mainly use detected emotion, while empathy also includes cognitive understanding of the user's situation and implied feelings.
Method
CEM leverages external commonsense knowledge to infer information about users' situations and feelings and use it to enhance empathetic response generation.
Results
CEM generates more informative and empathetic responses than previous methods in both automatic and manual evaluations.
Takeaways & Limitations
Commonsense can support cognitive understanding of users' situations alongside emotion detection in empathetic response generation.
Abstract
from arXiv · showhide
A key trait of daily conversations between individuals is the ability to express empathy towards others, and exploring ways to implement empathy is a crucial step towards human-like dialogue systems. Previous approaches on this topic mainly focus on detecting and utilizing the user's emotion for generating empathetic responses. However, since empathy includes both aspects of affection and cognition, we argue that in addition to identifying the user's emotion, cognitive understanding of the user's situation should also be considered. To this end, we propose a novel approach for empathetic response generation, which leverages commonsense to draw more information about the user's situation and uses this additional information to further enhance the empathy expression in generated responses. We evaluate our approach on EmpatheticDialogues, which is a widely-used benchmark dataset for empathetic response generation. Empirical results demonstrate that our approach outperforms the baseline models in both automatic and human evaluations and can generate more informative and empathetic responses.
1 Introduction
Empathetic dialogue requires both emotional responsiveness and cognitive understanding of the user's situation. CEM uses commonsense to infer additional situational information and improve the informativeness and empathy of generated responses.
- Prior empathetic dialogue methods mainly detect and use the user's emotion when generating responses.
- Empathy includes affective responses to experiences and cognitive understanding of situations and implied feelings.
- Commonsense reasoning can infer unstated information about users' emotions, experiences, and situations before response generation.
- CEM leverages external commonsense knowledge about reactions, intentions, and desires to enhance cognitive understanding and empathy expression.
- Automatic and manual evaluations report that CEM generates more informative and empathetic responses than previous methods.
2 Preliminaries
The paper frames empathetic response generation around affective and cognitive understanding, then uses ATOMIC and COMET to derive commonsense inferences for dialogue contexts. It evaluates listener responses on the 25k-conversation EmpatheticDialogues benchmark.
- Empathy combines affective experience of others' emotions with cognitive understanding of situations and implicit mental states.
- Existing empathetic dialogue systems largely focus on context emotion detection and give less attention to cognitive empathy.
- Users may leave situational information implicit, so commonsense connects explicit utterances with implied feelings and circumstances.
- ATOMIC supplies five used relations for events: xEffect, xReact, xIntent, xNeed, and xWant; xAttr is excluded because it judges personal attributes.
- A modified BART-based COMET model generates ATOMIC commonsense inferences for events, including events unseen in ordinary daily data.
- EmpatheticDialogues contains 25k multi-turn conversations, each labeled with one of 32 emotions, and the model generates the listener's next response.
3 Methodology
CEM builds on Transformer by encoding dialogue context, acquiring affective and cognitive commonsense inferences, refining and selecting knowledge, and generating responses. It also trains emotion classification and diversity-aware objectives to produce informative, empathetic responses rather than generic ones.
- 3.1–3.5 Model Architecture: CEM organizes its process into context encoding, knowledge acquisition, context refinement, knowledge selection, and response generation.The model is built upon the standard Transformer.
- 3.1 Context Encoding: The dialogue history is concatenated with a [CLS] token, whose final hidden representation represents the whole context while dialogue-state embeddings distinguish speakers.The resulting sequence embedding combines word, positional, and dialogue-state embeddings before context encoding.
- 3.2 Knowledge Acquisition: COMET supplies five commonsense relations: xReact captures affective state, while xWant, xNeed, xIntent, and xEffect represent cognitive information about the situation.Affective relations are represented by averaged hidden states, whereas cognitive relations use the [CLS] representation.
- 3.3–3.4 Context Refinement and Knowledge Selection: Relation representations are concatenated with context at the token level, separately encoded for affective and cognitive groups, then weighted and mixed into a combined contextual representation.Token-level concatenation fuses commonsense information within each word, and an MLP combines the relation-specific representations.
- 3.5–3.6 Response Generation and Training Objectives: The decoder generates responses token by token using the commonsense-refined context through modified cross-attention, while emotion classification is trained with cross-entropy loss.The model additionally uses negative log-likelihood for target responses and Frequency-Aware Cross-Entropy to penalize high-frequency tokens and reduce generic responses.
4 Experiments
CEM is evaluated against established empathetic dialogue baselines using automatic metrics, aspect-based human preferences, ablations, and case studies. Results indicate that commonsense-supported affective and cognitive reasoning improves response quality, diversity, emotion classification, informativeness, and empathy.
- Evaluation Setup: Perplexity, Dist-n, and emotion-classification accuracy are used to assess response quality, diversity, and emotion prediction.Lower PPL indicates higher confidence, while Dist-n measures the proportion of unique n-grams.
- Automatic Evaluation: CEM achieves the lowest perplexity, considerably higher Dist-n, and much higher emotion-classification accuracy than the baselines.The results associate commonsense knowledge with higher overall response quality, greater diversity, and improved emotion detection.
- Human Evaluation: CEM outperforms the baselines in coherence, empathy, and informativeness in the aspect-based pairwise preference test.The test compares responses within the same context across three explicitly defined aspects; 100 response pairs were annotated by three workers each.
- Ablation Studies: Removing affective or cognitive components reduces emotion-classification performance, while removing diversity loss lowers Dist-n scores.These ablations indicate that both affective and cognitive information contribute to emotion classification, and diversity loss contributes to response diversity.
- Case Study: Case studies show CEM using commonsense to infer unstated intentions, implied emotions, and safety-relevant context that baselines miss.Examples include interpreting “ready for a puppy” as wanting to obtain a dog, recognizing “bittersweet” feelings, and prioritizing health and safety after “I was not hit.”
5 Conclusions and Future Work
The paper concludes that commonsense knowledge can improve understanding of users’ situations and feelings in empathetic dialogue generation. It also suggests extending this direction to related tasks such as emotional support.
- Conclusion: CEM uses commonsense knowledge to improve understanding of users’ situations and feelings, producing more informative and empathetic responses.The conclusion reports support from both automatic and manual evaluations.
- Future Work: Future work can extend commonsense-based approaches to empathetic response generation and related tasks such as emotional support.