I like fish, especially dolphins: Addressing Contradictions in Dialogue Modeling

Yixin Nie, Mary Williamson, Mohit Bansal, Douwe Kiela, Jason Weston

Abstract

To quantify how well natural language understanding models can capture consistency in a general conversation, we introduce the DialoguE COntradiction DEtection task (DECODE) and a new conversational dataset containing both human-human and human-bot contradictory dialogues. We then compare a structured utterance-based approach of using pre-trained Transformer models for contradiction detection with the typical unstructured approach. Results reveal that: (i) our newly collected dataset is notably more effective at providing supervision for the dialogue contradiction detection task than existing NLI data including those aimed to cover the dialogue domain; (ii) the structured utterance-based approach is more robust and transferable on both analysis and out-of-distribution dialogues than its unstructured counterpart. We also show that our best contradiction detection model correlates well with human judgments and further provide evidence for its usage in both automatically evaluating and improving the consistency of state-of-the-art generative chatbots.

Introduction: The Consistency Problem

Modern open-domain chatbots, such as BlenderBot, have reached impressive levels of fluency. However, they frequently suffer from "consistency errors"—jarring self-contradictions that break the illusion of intelligence and erode user trust. For example, a bot might claim to be a high school Spanish teacher in one turn, only to admit a few turns later that it does not speak Spanish at all.

Examples of self-contradictions in BlenderBot 2.7B, where the model fails to maintain a consistent persona or plan.

These contradictions violate the Gricean maxim of quality: "do not say what you believe to be false." Addressing this requires moving beyond standard Natural Language Inference (NLI) datasets, which typically rely on static sentence pairs and fail to capture the nuances of multi-turn conversational flow.

The DECODE Dataset

To address the scarcity of dialogue-specific contradiction data, we introduce DECODE (DialoguE COntradiction DEtection). This large-scale resource includes:

  • Main Dataset: Approximately 35,000 dialogues featuring human-written contradictions where annotators deliberately contradicted previous turns.
  • Human-Bot Test Set: Real-world interactions between humans and state-of-the-art bots, containing over 8,000 utterances with human-labeled contradictions.
  • Checklist Sets: Specialized variants like A2T (Add Two Turns) and RCT (Remove Contradicting Turns) designed to test whether models truly understand logical consistency or are simply exploiting superficial patterns.

Modeling Strategies: Structured vs. Unstructured

We evaluated two primary architectures using Transformer models like RoBERTa and BART:

  1. Unstructured Approach: The model processes the entire concatenated dialogue history and the new response as a single string. It relies on self-attention to identify conflicts within the block.
  2. Structured Utterance-based Approach: This novel method pairs the latest utterance with each previous utterance from the same speaker individually. The final contradiction score is the maximum probability found across these pairs. This approach forces the model to pinpoint the exact source of the contradiction, providing inherent explainability.

Results and Insights

Our experiments demonstrate that dialogue contradiction detection is distinct from standard NLI. Models trained on DECODE significantly outperformed those trained on general NLI data (e.g., MNLI), achieving over 93% accuracy on the main test set.

Accuracy comparison between the structured (Utterance-based) and unstructured models across various test sets, highlighting the robustness of the structured approach.

A key finding is the "laziness" of unstructured models. In the RCT test—where the specific turn causing a contradiction is removed—unstructured models often continued to flag a contradiction based on superficial keywords. In contrast, the structured approach proved far more robust, correctly identifying when a contradiction was no longer present. Furthermore, the structured model showed superior transferability to out-of-distribution (OOD) human-bot dialogues, achieving an AUC of 87.16 compared to 73.40 for the unstructured version.

Improving Chatbot Consistency

Beyond detection, we applied our model to Generation Re-ranking. By using the contradiction detector to score and filter candidate responses during a chatbot's decoding phase, we reduced human-judged contradictions in BlenderBot from 31.8% to 25.6%.

Conclusion and Future Work

While DECODE provides a powerful tool for improving chatbot consistency, challenges remain. Current models primarily focus on self-contradictions rather than cross-speaker inconsistencies. Future research should explore integrating these detectors directly into the training process (e.g., via unlikelihood training) to prevent contradictions from being generated in the first place, rather than just filtering them post-hoc.

Summary updated

Abstract

To quantify how well natural language understanding models can capture consistency in a general conversation, we introduce the DialoguE COntradiction DEtection task (DECODE) and a new conversational dataset containing both human-human and human-bot contradictory dialogues. We then compare a structured utterance-based approach of using pre-trained Transformer models for contradiction detection with the typical unstructured approach. Results reveal that: (i) our newly collected dataset is notably more effective at providing supervision for the dialogue contradiction detection task than existing NLI data including those aimed to cover the dialogue domain; (ii) the structured utterance-based approach is more robust and transferable on both analysis and out-of-distribution dialogues than its unstructured counterpart. We also show that our best contradiction detection model correlates well with human judgments and further provide evidence for its usage in both automatically evaluating and improving the consistency of state-of-the-art generative chatbots.
BibTeX
@inproceedings{nie2021i,
  author = {Yixin Nie and Mary Williamson and Mohit Bansal and Douwe Kiela and Jason Weston},
  title = {{I like fish, especially dolphins: Addressing Contradictions in Dialogue Modeling}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
  url = {https://openreview.net/forum?id=50VID_9Ftv8},
}
Figure