Learning a Structural Causal Model for Intuition Reasoning in Conversation

Hang Chen, Bingyu Liao, Jing Luo, Wenjing Zhu, Xinyu Yang

Abstract

Reasoning, a crucial aspect of NLP research, has not been adequately addressed by prevailing models including Large Language Models. Conversation reasoning, as a critical component of it, remains largely unexplored due to the absence of a well-designed cognitive model. In this paper, inspired by intuition theory on conversation cognition, we develop a conversation cognitive model (CCM) that explains how each utterance receives and activates channels of information recursively. Besides, we algebraically transformed CCM into a structural causal model (SCM) under some mild assumptions, rendering it compatible with various causal discovery methods. We further propose a probabilistic implementation of the SCM for utterance-level relation reasoning. By leveraging variational inference, it explores substitutes for implicit causes, addresses the issue of their unobservability, and reconstructs the causal representations of utterances through the evidence lower bounds. Moreover, we constructed synthetic and simulated datasets incorporating implicit causes and complete cause labels, alleviating the current situation where all available datasets are implicit-causes-agnostic. Extensive experiments demonstrate that our proposed method significantly outperforms existing methods on synthetic, simulated, and real-world datasets. Finally, we analyze the performance of CCM under latent confounders and propose theoretical ideas for addressing this currently unresolved issue.

Introduction

While Large Language Models (LLMs) have demonstrated proficiency in "rational reasoning" tasks, such as step-by-step mathematical problem solving, they frequently struggle with intuitive reasoning—the social logic inherent in daily conversation. In a dialogue, an utterance is not merely a reaction to the preceding turn; it is shaped by unobservable "Mental States," including internal beliefs, emotions, and goals.

Current NLP approaches often treat conversation as a simple sequence or a correlation-based graph, failing to account for these "implicit" causes. Furthermore, the lack of datasets with labels for these hidden drivers hinders the development and evaluation of models capable of true conversational understanding. This paper addresses these gaps by formalizing a cognitive framework for dialogue.

Methodology

The proposed approach bridges the gap between psychological theory and computational modeling through a three-step process.

The Conversation Cognitive Model (CCM)

The authors propose a model where Perception (understanding previous turns) and Mental State (internal drivers) combine to form a Plan, which then manifests as an Utterance or Action.

Causal Transformation (SCM)

To render this theory computable, the CCM is simplified into a Structural Causal Model (SCM). In this framework, utterances are treated as endogenous variables (observable effects), while mental states are modeled as exogenous variables (implicit causes or "noise"). This transformation allows the relationship between dialogue turns to be mathematically defined.

Transformation from the psychological CCM to a computable SCM with underlying equations.

Probabilistic Implementation

Since mental states are unobservable, the model employs Variational Inference to learn latent representations:

  • Encoder: A Graph Neural Network (GNN) with an attention mechanism predicts the distribution of implicit causes from input text.
  • Decoder: Reconstructs the causal representation of the utterance using predicted implicit causes and a causal strength matrix.
  • Optimization: The model is trained using the Evidence Lower Bound (ELBO), enabling it to learn hidden causes without explicit labels.

Experimental Setup

The study utilizes three distinct types of data to evaluate the model:

  1. RECCON: A real-world dataset for emotion-cause recognition in conversations, containing partial labels for emotional triggers.
  2. Synthetic Dataset: A benchmark created using Causal Additive Models (CAMs) following a strict Directed Acyclic Graph (DAG) structure, providing ground truth for both explicit and implicit causes.
  3. Simulation Dataset: Dialogues generated by GPT-4 based on specific "causal skeletons" (e.g., chains vs. complex multi-turn influences) to test the model's ability to distinguish causal flow structures.

Results and Discussion

The proposed model consistently outperformed baselines, including RoBERTa, GPT-3.5, GPT-4, and specialized GNN models like DAG-ERC and CAE.

Performance comparison showing the proposed model outperforming LLMs and GNN baselines across datasets.

Key Findings

  • Explicit Cause Extraction (ECE): On the RECCON dataset, the model achieved an F1 score of 77.59, significantly higher than the 73.17 of the best baseline and the 51.1 achieved by GPT-4.
  • Causal Discriminability: The SCM approach excels in "Reversal" tests—distinguishing if turn A caused B or vice versa—where LLMs often fail due to high semantic overlap.
  • Implicit Cause Recovery: High F1 scores in Implicit Cause Extraction (ICE) tasks (>96) suggest that the model's latent variables effectively capture the speakers' underlying emotional sentiments.

Conclusion

By algebraizing psychological intuition into a structural causal model, this work moves beyond "black-box" sequence modeling toward a framework where conversational relationships are mathematically grounded. While the model shows superior performance, latent confounders—unobserved variables influencing both cause and effect—remain a challenge. Future work will focus on expanding causal skeletons and refining manual verification of AI-generated dialogues to better capture human social nuances.

Summary updated

Abstract

Reasoning, a crucial aspect of NLP research, has not been adequately addressed by prevailing models including Large Language Model. Conversation reasoning, as a critical component of it, remains largely unexplored due to the absence of a well-designed cognitive model. In this paper, inspired by intuition theory on conversation cognition, we develop a conversation cognitive model (CCM) that explains how each utterance receives and activates channels of information recursively. Besides, we algebraically transformed CCM into a structural causal model (SCM) under some mild assumptions, rendering it compatible with various causal discovery methods. We further propose a probabilistic implementation of the SCM for utterance-level relation reasoning. By leveraging variational inference, it explores substitutes for implicit causes, addresses the issue of their unobservability, and reconstructs the causal representations of utterances through the evidence lower bounds. Moreover, we constructed synthetic and simulated datasets incorporating implicit causes and complete cause labels, alleviating the current situation where all available datasets are implicit-causes-agnostic. Extensive experiments demonstrate that our proposed method significantly outperforms existing methods on synthetic, simulated, and real-world datasets. Finally, we analyze the performance of CCM under latent confounders and propose theoretical ideas for addressing this currently unresolved issue.
BibTeX
@article{chen2024learning,
  author = {Hang Chen and Bingyu Liao and Jing Luo and Wenjing Zhu and Xinyu Yang},
  title = {{Learning a Structural Causal Model for Intuition Reasoning in Conversation}},
  journal = {IEEE Trans. Knowl. Data Eng.},
  year = {2024},
  url = {https://openreview.net/forum?id=adRmgE6UmZ},
}
Figure