PaperlayerView exact revision on arXiv

Source-linked AI summary

CARE: Causality Reasoning for Empathetic Responses by Conditional Graph Generation

Jiashuo Wang, Yi Cheng, Wenjie Li

arXiv:2211.00255v2cs.CL

TL;DR

Prior empathetic response methods omit links between user experiences and reason causalities independently. CARE uses CVGAE to reason plausible causalities jointly and infuses them into response generation, achieving state-of-the-art performance on empathetic response generation.

  • Problem

    Prior methods consider emotion-to-experience causalities but omit links between experiences and reason causalities independently.

    Source: §1 · p. 1

  • Method

    CARE uses CVGAE to generate a causal graph conditioned on emotion, dialogue history, and predicted future content, then infuses causalities into response generation with decoder multi-source attention.

    Source: §1 · p. 2

  • Results

    CARE achieves state-of-the-art performance on empathetic response generation on the EMPATHETICDIALOGUES benchmark.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors report that CARE improves the model’s understanding of users’ feelings and experiences.

    Source: §1 · p. 2

    The discussion leaves integration of sentiment and commonsense knowledge with empathy as a problem for further exploration.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Recent approaches to empathetic response generation incorporate emotion causalities to enhance comprehension of both the user's feelings and experiences. However, these approaches suffer from two critical issues. First, they only consider causalities between the user's emotion and the user's experiences, and ignore those between the user's experiences. Second, they neglect interdependence among causalities and reason them independently. To solve the above problems, we expect to reason all plausible causalities interdependently and simultaneously, given the user's emotion, dialogue history, and future dialogue content. Then, we infuse these causalities into response generation for empathetic responses. Specifically, we design a new model, i.e., the Conditional Variational Graph Auto-Encoder (CVGAE), for the causality reasoning, and adopt a multi-source attention mechanism in the decoder for the causality infusion. We name the whole framework as CARE, abbreviated for CAusality Reasoning for Empathetic conversation. Experimental results indicate that our method achieves state-of-the-art performance.

1 Introduction

CARE addresses gaps in prior causality-aware empathy models by reasoning across plausible causalities jointly, then infusing them into empathetic response generation.

1 Introduction

  • The paper frames empathy as involving both affective emotional understanding and reaction, and cognitive comprehension of another person’s experiences and situation.Source: §1 · p. 1
  • Earlier methods considered emotion-to-experience causalities but omitted causal links between experiences and reasoned links independently.The authors argue that experience-to-experience links can clarify causal relationships that are not apparent in isolation.Source: §1 · pp. 1–2
  • CARE reasons plausible causalities jointly and simultaneously, including links stated or implied in dialogue history and predicted future dialogue.The generated causal graph is conditioned on the user’s emotion, dialogue history, and predicted future dialogue content.Source: §1 · p. 2
  • CARE infuses reasoned causalities into response generation through decoder multi-source attention and integrates reasoning and generation with multitask learning.Source: §1 · p. 2
  • CARE achieves state-of-the-art performance on empathetic response generation on the EMPATHETICDIALOGUES benchmark.The authors also report improved understanding of users’ feelings and experiences.Source: §1 · p. 2
  • CVGAE formulates causality reasoning as conditional graph generation and uses latent variables for conditional structure prediction.The graph represents feelings and experiences as word nodes and causal relationships as edges.Source: §1 · p. 2

2 Related Work

Related work moves from methods centered on affect toward approaches incorporating emotion causes and effects, while the latter still reason causalities independently.

2 Related Work

  • Earlier methods emphasize affect: MoEL combines emotion-specific decoder outputs, MIME models polarity and emotion variation, and Li et al. (2020a) capture token-level emotion nuances.Li et al. (2020a) also use adversarial learning to incorporate user feedback.Source: §2 · p. 2
  • Later approaches incorporate cognition through emotion causes or effects: Wang et al. (2021), Gao et al. (2021), and Kim et al. (2021) identify causes, while CEM (Sabour et al., 2022) generates causes and effects.Source: §2 · p. 2
  • The reviewed cause-and-effect methods obtain causalities independently, including classifier-based cause detection and CEM’s context refinement.Source: §2 · p. 2

3 Preliminary

The section introduces Transformer-based response generation and the Variational Graph Auto-Encoder (VGAE), which provides the graph modeling basis for the paper’s later method.

3.1 Transformer-based Response Generation

  • The Transformer encoder represents dialogue context, while its decoder generates responses using self-attention, context cross-attention, and feed-forward sublayers.A final linear transformation and softmax predict token probabilities.Source: §3.1 · p. 3

3.2 Variational Graph Auto-Encoder

  • VGAE infers graph latent variables from an undirected graph and reconstructs its adjacency matrix using an inner product between latent variables.The inference model uses a recognition net conditioned on the graph and adjacency matrix.Source: §3.2 · p. 3

4 Method

CARE constructs causal graphs and uses CVGAE to reason about plausible causalities, then incorporates the graph into response generation through decoder attention. Its training objective jointly integrates causality reasoning and response generation.

4 Method

  • CARE generates a causal graph to reason about plausible causalities interdependently, then infuses the graph into response generation through decoder multi-source attention.The framework uses CVGAE for graph generation.Source: §4 · pp. 3–4

4.1 Graph Construction

  • The prior graph contains causal relations from previous user utterances, while the posterior graph also includes relations involving the next user utterance.Both graphs use a shared node set drawn from emotion words, words in previous user utterances, and their one-hop neighbors in the Cause Effect Graph.Source: §4.1 · p. 4

4.2 Conditional Variational Graph Auto-Encoder (CVGAE)

  • CVGAE conditions graph-latent inference on dialogue context, user emotion, and a future-dialogue latent, then reconstructs causal relations and selects top-k by probability.The context latent is generated from the dialogue context and golden response during training.Source: §4.2 · pp. 4–5
  • At inference, CVGAE uses a prior net to approximate the recognition distribution because the golden response and posterior adjacency matrix are unavailable.The prior net is conditioned on the prior graph and context.Source: §4.2 · p. 5

4.3 Graph-Infused Response Generation

  • The modified decoder cross-attention attends to both dialogue context and the causal graph to generate the final response.The remaining original decoder components are unchanged.Source: §4.3 · p. 5

4.4 Training Objective

  • Multi-task training combines causality reasoning, evaluated by graph reconstruction and posterior-prior similarity, with response generation loss.CARE is trained by maximizing the sum of the response-generation and graph losses.Source: §4.4 · p. 5

5 Experiments

Experiments on EmpatheticDialogues compare CARE with affective-only and affection-plus-cognition systems. CARE leads the reported automatic metrics and most human empathy and relevance ratings, while ablations, relationship-count analysis, and a case study examine its behavior.

5.1 Dataset

  • The experiments use EmpatheticDialogues, with 25,000 crowdsourced one-on-one conversations across 32 balanced emotion categories and an approximately 80/10/10 train-validation-test split.Source: §5.1 · p. 5

5.2 Comparison Models

  • Baselines included three affective-only models and four models considering affection and cognition; examples include MoEL and CEM.Source: §5.2 · p. 6

5.3 Implementation Details

  • CARE uses 300-dimensional pretrained GloVe embeddings, 800 nodes, and 512 selected relationships, and comparison models use official code where available.The models share embeddings across the encoder, CVGAE, and decoder; GEEMIME was implemented from official instructions because only test code and instructions were provided.Source: §5.3 · p. 6

5.4 Automatic Evaluation

  • CARE outperforms the comparison models on all reported automatic metrics; its perplexity score is unavailable for GEEMIME because PPL is less relevant to that model's structure.The authors report high BLEU and BERTScore as indicating more human-like responses; they also describe the advantages as statistically significant and stable.Source: §5.4 · pp. 6–7, Table 1 · p. 7

5.5 Human Ratings

  • CARE achieves the highest Emp. and Rel. scores compared with most prior models and relatively high Flu., while models considering both empathy aspects do not always outperform affective-only models.The human ratings use a 5-point scale, with 5 as the best score.Source: §5.5 · p. 7, Table 2 · p. 7

5.6 Model Analysis

  • Both ablations surpass prior models on automatic metrics but score relatively low in human evaluation; full CARE scores higher than the no-condition variant on empathy and relevance.The no-condition variant scores higher than the no-reasoning variant on Emp. and Rel.; the authors attribute this comparison to reasoning plausible causalities not mentioned in dialogue history.Source: §5.6 · p. 8
  • As the number of relationships infused increases, BLEU-4 first rises and then falls, which the authors associate with sufficient causalities helping and excess ones adding noise.Figure 3 reports averages over five runs for each k.Source: §5.6 · p. 8, Figure 3 · p. 8

5.7 Case Study

  • In the case study, CARE responds more empathically and shows comprehension of the user's feelings and experiences, including apprehension linked to low confidence and recognition of a proposed solution.The example response recognizes the user's “great idea” as an effective solution.Source: §5.7 · p. 8, Table 5 · p. 9

6 Conclusion

CARE reasons plausible conversational causalities jointly to understand users’ feelings and experiences, then incorporates them into empathetic response generation. The authors report that it outperforms prior methods in automatic and manual evaluations.

6 Conclusion

  • CARE frames conversational causality reasoning as generating plausible causalities interdependently and simultaneously.The stated aim is deeper understanding of users’ feelings and experiences in empathetic dialogue.Source: §6 · p. 9
  • CVGAE uses latent variables for conditional structure prediction and predicted future conversation content to reason about causalities.Source: §6 · p. 9
  • A decoder multi-source attention mechanism infuses reasoned causalities into response generation, and CARE outperforms prior methods in automatic and manual evaluations.Source: §6 · p. 9

Limitations

The paper focuses on causal knowledge to improve empathy, especially cognition, while noting that other knowledge may also contribute. It identifies connecting knowledge types to empathy and fusing them as open questions.

Limitations

  • The paper targets empathy’s affection and cognition, especially cognition, by incorporating reasoned causal knowledge into response generation.Source: Limitations · p. 9
  • Sentiment and commonsense knowledge may also contribute to affection and cognition, beyond the causal knowledge incorporated here.Source: Limitations · p. 9
  • The authors identify connecting empathy with different knowledge types and fusing multiple kinds of knowledge in one model as worthwhile future problems.Source: Limitations · p. 9

Ethical Considerations

The experiment uses the open-sourced EMPATHETICDIALOGUES benchmark, collected through MTurk interaction. The authors state that privacy is protected and no personal information is included.

Ethical Considerations

  • The experiment uses the widely used, open-sourced EMPATHETICDIALOGUES benchmark collected through interaction with Amazon Mechanical Turk.Source: Ethical Considerations · p. 9
  • The authors report that user privacy is protected and the dataset contains no personal information.Source: Ethical Considerations · p. 9

A.1 Automatic Evaluation

Automatic evaluation uses modified dialogue-evaluation code from Csáky et al. (2019), and the full results are presented in Table 6.

A.1 Automatic Evaluation

A.2 Human Evaluation

The human-rating system shows annotators prior dialogue turns and the user’s emotion, while responses from different models are presented in random order to conceal their sources. Ratings use reference statements and examples for empathy, relevance, and fluency, with annotators encouraged to compare responses before grading.

A.2 Human Evaluation

  • Annotators see prior dialogue turns and the user’s emotion, while model responses appear in random order so their sources cannot be distinguished.Figure 4 shows the rating interface.Source: §A.2 · p. 11, Figure 4 · p. 13
  • Human ratings use reference examples to assess empathy, relevance, and fluency, including emotional care, topic fit, repetition, grammar, and conflicting content.For example, the guidelines prefer asking whether the user called the police over an unrelated movie question after a theft account.Source: §A.2 · p. 11
  • Annotators are encouraged to compare different responses mentally before grading each one.Source: §A.2 · p. 11
Loading 2211.00255v2…