Rethinking Dialogue State Tracking with Reasoning

wenqiang leiarXiv 2020 · arXiv

TL;DR

This paper introduces ReDST, a novel dialogue state tracking (DST) approach that moves beyond simply classifying slot-value pairs. ReDST leverages both turn-level reasoning and a bipartite graph representation of the backend database to achieve significantly improved accuracy, particularly for complex entities like names.

What Problem Does This Paper Address?

Dialogue State Tracking (DST) is a crucial component in dialogue management systems, responsible for understanding and maintaining a user's goals throughout a conversation. Traditional DST methods often treat this as a classification problem, mapping utterances to predefined slot-value pairs. While effective for simple queries, these methods struggle with:

  • Long-range dependencies: Understanding how information from earlier turns influences current goals.
  • Reasoning capabilities: Inferring user intent when explicit information is missing or requires combining multiple pieces of information.
  • Backend database integration: Effectively using external knowledge bases to validate or infer user requirements.

The paper argues that current DST approaches are too reliant on domain ontologies and lack the ability to reason over dialogue history and backend data, leading to limitations in handling complex user intents and goal shifts.

Methods

ReDST proposes a three-component architecture:

  1. Turn Belief Generator: This component uses a fine-tuned BERT model to predict the belief state for a single turn. It employs a multi-task learning setup, combining sequence classification (for slot values like "yes," "no," "don't care") and token classification (for extracting specific slot value spans). This leverages BERT's powerful contextual understanding for more accurate turn-level predictions.
  2. Bipartite Graph Belief Propagator: This novel component abstracts the backend database into a bipartite graph where one set of nodes represents entities (e.g., restaurants) and the other represents attributes (e.g., price range, cuisine). Beliefs from the current turn and the previous joint belief are mapped onto this graph. Belief propagation is then performed to infer credibility scores for slot values, effectively reasoning about how database information relates to user requirements. This helps identify potential conflicts or validate user requests.
  3. Incremental Belief Generator: This component integrates the turn belief, the previous joint belief, and the propagated beliefs from the bipartite graph to incrementally infer the current joint belief. The paper explores different integration methods, including simple rules, MLPs, and GRU cells, finding GRU to be most effective due to its ability to adaptively remember and forget information.

Datasets

The paper evaluates its method on the MultiWOZ 2.1 dataset. This is a large-scale, multi-domain dialogue dataset that has been corrected for noisy annotations compared to its predecessor, MultiWOZ 2.0. The dataset covers seven distinct domains and contains over 10,000 dialogues.

Key Insights and Ideas

  • Reasoning is essential: The paper strongly advocates for moving beyond simple classification and incorporating reasoning over both dialogue turns and backend data for robust DST.
  • Bipartite graph for database reasoning: Representing the backend database as a bipartite graph and using belief propagation is an effective way to leverage external knowledge for DST, especially for identifying relationships between entities and their attributes.
  • BERT as a strong foundation: Fine-tuning BERT for turn belief generation provides a powerful starting point, significantly improving accuracy over previous methods that relied on simpler RNNs.
  • Incremental inference is key: The recursive nature of dialogue necessitates an incremental approach to updating the joint belief, where each turn builds upon the previous state and new information.

Key Results

ReDST, particularly the ReDST_GRU variant, significantly outperforms state-of-the-art methods on the MultiWOZ 2.1 dataset.

  • Joint Accuracy Improvement: ReDST_GRU achieved a 38.6% improvement in joint accuracy over the previous state-of-the-art TRADE model.
  • Slot Error Rate Reduction: The model showed substantial improvements in error rates for name slots (restaurant, hotel, attraction), dropping by an average of 82.02%. This highlights the effectiveness of the bipartite graph reasoning for complex entities.
  • Turn Belief Accuracy: Even the turn belief generator component (without reasoning) showed a 4.2% improvement in joint accuracy over single-turn TRADE, demonstrating the benefit of BERT.

Limitations and Future Work

  • Database Reasoning Integration: While the bipartite graph approach is promising, the paper notes that integrating database reasoning with rule-based methods can be challenging, and for some models (like ReDST_MLP), it even led to worse performance, suggesting that finding the right integration strategy is crucial.
  • Simple Belief Propagation: The belief propagation method used is described as simple. The authors suggest exploring more advanced graph reasoning algorithms.
  • Future Directions: The authors plan to explore more advanced reasoning algorithms for turns and graphs, investigate more detailed relationships between nodes in the graph, and examine other possibilities for incremental reasoning to achieve even more accurate user intention summarization.

Summary updated

Abstract

Tracking dialogue states to better interpret user goals and feed downstream policy learning is a bottleneck in dialogue management. Common practice has been to treat it as a problem of classifying dialogue content into a set of pre-defined slot-value pairs, or generating values for different slots given the dialogue history. Both have limitations on considering dependencies that occur on dialogues, and are lacking of reasoning capabilities. This paper proposes to track dialogue states gradually with reasoning over dialogue turns with the help of the back-end data. Empirical results demonstrate that our method significantly outperforms the state-of-the-art methods by 38.6% in terms of joint belief accuracy for MultiWOZ 2.1, a large-scale human-human dialogue dataset across multiple domains.
BibTeX
@misc{lei2020rethinking,
  author = {wenqiang lei},
  title = {{Rethinking Dialogue State Tracking with Reasoning}},
  year = {2020},
  eprint = {2005.13129},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.13129},
}
Figure