Do Neural Dialog Systems Use the Conversation History Effectively? An Empirical Study

Chinnadhurai Sankar, Sandeep Subramanian, Christopher J. Pal, Sarath Chandar, Yoshua Bengio

Abstract

Neural generative models have become increasingly popular when building conversational agents. They offer flexibility, can be easily adapted to new domains, and require minimal domain engineering. A common criticism of these systems is that they seldom understand or use the available dialog history effectively. In this paper, we take an empirical approach to understanding how these models use the available dialog history by studying the sensitivity of the models to artificially introduced unnatural changes or perturbations to their context at test time. We experiment with 10 different types of perturbations on 4 multi-turn dialog datasets and find that commonly used neural dialog architectures like recurrent and transformer-based seq2seq models are rarely sensitive to most perturbations such as missing or reordering utterances, shuffling words, etc. Also, by open-sourcing our code, we believe that it will serve as a useful diagnostic tool for evaluating dialog systems in the future.

Introduction

Generative neural dialog systems, such as sequence-to-sequence (seq2seq) and Transformer models, are the foundation of modern conversational AI. Despite their ability to produce fluent responses, they are often criticized for generating generic, repetitive content and failing to maintain topical consistency over multiple turns. This has led to the hypothesis that these models do not effectively utilize the provided conversation history. This study provides empirical evidence for this claim by systematically perturbing the dialog history at test time and measuring the resulting impact on model performance.

Methodology

The authors evaluate three standard architectures: a vanilla LSTM seq2seq, an LSTM seq2seq with attention, and a Transformer. These models are tested on four diverse datasets: bAbI dialog (Task 5), Persona Chat, Dailydialog, and MutualFriends.

To probe the models' reliance on context, ten synthetic perturbations are introduced to the dialog history exclusively at test time. The underlying logic is that if a model ignores certain information, its performance—measured by perplexity (PPL)—will remain stable even when that information is corrupted. Perturbations are categorized into:

  1. Utterance-level: Shuffling or reversing the order of utterances, dropping the first or last utterance, or truncating the history to the $k$ most recent turns.
  2. Word-level: Randomly shuffling or reversing word order within utterances, dropping 30% of words, or removing all nouns or verbs.

Key Results

The experimental results reveal a startling lack of sensitivity to history corruption. In many cases, extreme changes to the input history resulted in negligible increases in perplexity.

Table 1: Perplexity changes (ΔPPL) across different models and datasets under various utterance-level and word-level perturbations.

As shown in Table 1, the Transformer model on the Dailydialog dataset exhibited a perplexity increase of only 0.15 when words were completely shuffled, despite a baseline perplexity of 28.73. Across the 40 test cases, the LSTM with attention proved to be the most sensitive model, while the Transformer was generally the least sensitive to perturbations.

Analysis of Context Usage

A key finding is that models rely heavily on the most recent utterance while largely ignoring earlier context. This is most evident when examining the effect of truncating the dialog history.

Figure 1: Perplexity change as a function of truncated dialog history length across two datasets.

Figure 1 illustrates that perplexity remains relatively flat as the history is truncated from six turns down to two. A significant spike in perplexity only occurs when the history is reduced to a single turn, suggesting that the models primarily function as single-turn response generators.

Furthermore, the Transformer's high insensitivity to word-reordering suggests it may be learning bag-of-words-like representations rather than capturing complex syntax or conversational dynamics. In contrast, recurrent models with attention mechanisms showed slightly higher sensitivity to utterance ordering, indicating a better, albeit still limited, grasp of conversational flow.

Conclusion

This empirical study demonstrates that current neural dialog models often act as shallow pattern matchers rather than sophisticated conversationalists. They are remarkably robust to unnatural changes in history, which implies they are not truly "understanding" the multi-turn context. The authors have open-sourced their perturbation framework to help the community diagnose and improve the context-awareness of future dialog systems.

Summary updated

Abstract

Neural generative models have been become increasingly popular when building conversational agents. They offer flexibility, can be easily adapted to new domains, and require minimal domain engineering. A common criticism of these systems is that they seldom understand or use the available dialog history effectively. In this paper, we take an empirical approach to understanding how these models use the available dialog history by studying the sensitivity of the models to artificially introduced unnatural changes or perturbations to their context at test time. We experiment with 10 different types of perturbations on 4 multi-turn dialog datasets and find that commonly used neural dialog architectures like recurrent and transformer-based seq2seq models are rarely sensitive to most perturbations such as missing or reordering utterances, shuffling words, etc. Also, by open-sourcing our code, we believe that it will serve as a useful diagnostic tool for evaluating dialog systems in the future.
BibTeX
@inproceedings{sankar2019do,
  author = {Chinnadhurai Sankar and Sandeep Subramanian and Christopher J. Pal and Sarath Chandar and Yoshua Bengio},
  title = {{Do Neural Dialog Systems Use the Conversation History Effectively? An Empirical Study}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
  doi = {10.18653/v1/p19-1004},
}
Figure