Source-linked AI summary
Do Neural Dialog Systems Use the Conversation History Effectively? An Empirical Study
Chinnadhurai Sankar, Sandeep Subramanian, Christopher Pal, Sarath Chandar, Yoshua Bengio
TL;DR
Neural dialog systems are criticized for using conversation history weakly, but this concern has received limited empirical investigation. The paper perturbs dialog histories at test time and measures recurrent and transformer seq2seq sensitivity, finding minimal effects from many drastic changes. These patterns motivate perturbation analysis as a diagnostic tool for understanding dialog-system behavior.
Problem
Neural dialog models can generate fluent responses but may not effectively understand or use dialog history, and this concern has had little empirical investigation.
Method
The study artificially perturbs conversation histories at test time and measures changes in per-token perplexity for recurrent and transformer-based seq2seq models.
Results
Both recurrent and transformer-based seq2seq models show little sensitivity to most perturbations, including drastic changes, while recurrent models are more sensitive to utterance ordering than transformers.
Takeaways & Limitations
Perturbing dialog-history structure can serve as a useful diagnostic tool for evaluating model behavior and understanding which information models use.
Takeaways & Limitations
The evaluated models may not be state-of-the-art on the selected datasets, although the authors consider them competitive and commonly used as baselines.
Abstract
from arXiv · showhide
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.
1 Introduction
Neural dialog models can generate fluent responses but have been criticized for not using dialog history effectively, a claim with little empirical validation. This work tests that claim through synthetic history perturbations and finds broad insensitivity, with recurrent models more responsive to utterance ordering than transformers.
- 1 Introduction: The work addresses a gap in empirical evidence about whether neural dialog systems actually use the history available during response generation.Prior criticism attributed incoherent or repetitive responses to history use, but the paper frames this as insufficiently tested.
- 1 Introduction: Synthetic perturbations reveal that recurrent and transformer-based seq2seq models are insensitive to most dialog-history changes.The study evaluates how model behavior changes when available context is artificially altered.
- 1 Introduction: Both architectures remain particularly insensitive when words are randomly shuffled or reversed within every history utterance.These extreme perturbations directly disrupt word order while preserving the utterances’ content inventory.
- 1 Introduction: Recurrent models are more sensitive than transformers to utterance ordering, suggesting better modeling of conversation dynamics.The architectural difference concerns ordering across utterances rather than general response fluency.
2 Related Work
Related work examines how neural models use context under perturbations and how sequence-model architectures encode structure. These studies motivate testing dialog-history use across recurrent, transformer, and attention-based designs.
- 2 Related Work: This paper extends perturbation-based behavioral analysis to generative dialog systems and compares architectural and attention effects on context use.The stated focus is whether recurrent versus transformer architectures and attention mechanisms change sensitivity to available dialog information.
- 2 Related Work: Prior analyses found unconditional language models use mainly nearby context, whereas machine-translation models can be affected by noise anywhere in the input.The contrast motivates examining how contextual sensitivity varies across conditional and unconditional modeling settings.
- 2 Related Work: Research on inductive biases reports that recurrent models better capture hierarchical structure, while transformers and convolutional models do not outperform RNNs on long-distance agreement.Transformers nevertheless excel at word-sense disambiguation, showing that architectural strengths depend on the structure being modeled.
3 Experimental Setup
The study evaluates recurrent and transformer-based seq2seq dialog models by measuring perplexity changes after isolated perturbations to conversation histories across four multi-turn datasets.
- Experimental Design: Models are evaluated by the increase in per-token perplexity after test-time perturbations to their conversation histories.All models are trained without perturbations; sensitivity is measured only at test time.
- Datasets: The study uses four multi-turn dialog datasets spanning synthetic goal-oriented, open-domain chit-chat, and multi-topic conversational settings.The datasets include bAbI dialog, Persona Chat, Dailydialog, and MutualFriends.
- Types of Perturbations: Utterance-level perturbations shuffle, reverse, drop, or truncate conversation-history utterances, including retaining only the k most recent utterances.Truncation uses k ≤ n, where n is the history length.
- Types of Perturbations: Word-level perturbations shuffle or reverse words and remove either 30% of words, all nouns, or all verbs within utterances.All perturbations are applied in isolation.
- Models: The experiments compare recurrent LSTM seq2seq models, including an attention decoder, with a transformer seq2seq model using ParlAI defaults.The recurrent encoder and decoder use two LSTM layers with 128-dimensional hidden states; the transformer uses 300-dimensional embeddings and hidden states, two layers, and two attention heads.
- Evaluation: Table 2 reports baseline test perplexity and perturbation-induced increases across models and datasets, with means and standard deviations over five runs.The table also includes an Only Last condition and marks the most sensitive model for each dataset-perturbation pair.
4 Results & Discussion
Across models, drastic dialog-history perturbations usually cause only tiny perplexity changes, revealing limited use of available context. Architecture still matters: attention increases reliance on earlier utterances, while transformers are less sensitive to conversational structure than recurrent models.
- Models show only tiny perplexity changes under most extreme dialog-history perturbations, suggesting they use far from all available information.
- Transformers are insensitive to word reordering, indicating they could learn bag-of-words-like representations.
- Attention-based seq2seq LSTMs and transformers use more earlier conversational context than vanilla seq2seq models when only the last utterance remains.
- Although transformers converge faster and reach lower test perplexities, they are less sensitive than recurrent models to perturbations scrambling cross-utterance structure.
5 Conclusion
The study probes generative neural dialog systems with synthetic history perturbations and finds recurrent and transformer seq2seq models largely unaffected by drastic changes. The authors propose this perturbation-based paradigm and open-source code as diagnostic tools for dialog systems and datasets.
- Synthetic perturbations reveal that recurrent and transformer-based seq2seq models are often not significantly affected by drastic dialog-history modifications.
- The perturbation paradigm and open-source code are intended as useful diagnostic tools for evaluating dialog systems and understanding information use in new datasets.