PaperlayerView exact revision on arXiv

Source-linked AI summary

MinTL: Minimalist Transfer Learning for Task-Oriented Dialogue Systems

Zhaojiang Lin, Andrea Madotto, Genta Indra Winata, Pascale Fung

arXiv:2009.12005v2cs.CLcs.AI

TL;DR

Task-oriented dialogue systems require complex design and substantial annotated data, while adapting pre-trained language models to this setting is nontrivial. MinTL transfers pre-trained Seq2Seq models, jointly learns state tracking and response generation, and uses Lev for compact state updates; on MultiWOZ, it improves state-of-the-art performance and remains competitive with 20% of the training data.

  • Problem

    Task-oriented dialogue systems require complex design and substantial human-annotated data, and adapting pre-trained language models to them is nontrivial.

  • Method

    MinTL is a plug-and-play framework using pre-trained Seq2Seq backbones to jointly learn dialogue state tracking and response generation, with Lev modeling state differences.

  • Results

    MinTL-based systems improve state-of-the-art performance, and achieve competitive results using 20% training data.

  • Takeaways & Limitations

    MinTL provides a simpler transfer-learning approach for task-oriented dialogue while retaining strong benchmark performance in full-data and low-resource settings.

Abstract

from arXiv · show

In this paper, we propose Minimalist Transfer Learning (MinTL) to simplify the system design process of task-oriented dialogue systems and alleviate the over-dependency on annotated data. MinTL is a simple yet effective transfer learning framework, which allows us to plug-and-play pre-trained seq2seq models, and jointly learn dialogue state tracking and dialogue response generation. Unlike previous approaches, which use a copy mechanism to "carryover" the old dialogue states to the new one, we introduce Levenshtein belief spans (Lev), that allows efficient dialogue state tracking with a minimal generation length. We instantiate our learning framework with two pre-trained backbones: T5 and BART, and evaluate them on MultiWOZ. Extensive experiments demonstrate that: 1) our systems establish new state-of-the-art results on end-to-end response generation, 2) MinTL-based systems are more robust than baseline methods in the low resource setting, and they achieve competitive results with only 20\% training data, and 3) Lev greatly improves the inference efficiency.

1 Introduction

MinTL simplifies task-oriented dialogue system design by transferring pre-trained Seq2Seq models and jointly learning dialogue state tracking and response generation. Its Lev representation supports efficient state updates, while experiments show strong full-data and low-resource performance.

  • Framework: MinTL is a plug-and-play transfer learning framework that jointly learns dialogue state tracking and response generation with pre-trained Seq2Seq models.The framework is instantiated with T5 and BART backbones and avoids ad hoc task-specific modules.
  • Framework: Lev models the difference between old and new dialogue states instead of carrying over complete previous states.The updated state is then used for knowledge-base search before response decoding.
  • Efficiency: Lev enables minimal-length dialogue-state generation and greatly reduces inference latency.The representation is designed to edit the previous dialogue state efficiently.
  • Results: Two pre-trained backbones improve state-of-the-art results by a large margin on task-oriented dialogue benchmarks.Experiments evaluate T5 and BART on MultiWOZ across dialogue-state tracking and end-to-end response generation.
  • Results: Using 20% of the training data, MinTL-based systems achieve competitive results compared with the state of the art.This demonstrates robustness in a low-resource setting with fewer human annotations.

2 Related Work

Prior task-oriented dialogue systems commonly rely on pipelines, task-specific designs, and extensive annotations. Related work on pre-trained language models motivates a simpler Seq2Seq transfer framework that can retain performance with less data.

  • Pre-trained Language Models: Pre-trained language models have improved many NLP applications, but adapting them to task-oriented dialogue remains nontrivial.Seq2Seq models address the tension between bidirectional understanding and generative response production.
  • Task-Oriented Dialogue: Task-oriented dialogue systems traditionally use pipelines for belief tracking, dialogue management, and response generation.End-to-end systems were introduced to simplify this design, but earlier approaches still used varied methodologies.
  • Task-Oriented Dialogue: Lev edits dialogue states through insertion, deletion, and substitution operations rather than reproducing the entire state.The figure associates these operations with blue, red, and green updates, respectively.
  • Task-Oriented Dialogue: State-of-the-art systems achieved strong MultiWOZ results but relied on task-specific design and extensive human annotations.These requirements motivate transfer learning with pre-trained Seq2Seq models and smaller training sets.

3 Methodology

The methodology formalizes dialogue contexts and states for task-oriented dialogue tracking. It frames tracking as updating slot values over turns, while Lev supplies a compact edit representation for those updates.

  • Notations: The methodology defines a dialogue as alternating user utterances and system responses, with a context window determining the current dialogue context.The dialogue state at each turn is represented separately from the context and response.
  • Notations: Each dialogue state is a dictionary mapping domain-slot pairs to slot values.An absent domain-slot key is represented by the empty string ε.
  • Dialogue State Tracking: The goal of dialogue state tracking is to track slot values for each domain mentioned in the dialogue.The methodology introduces Lev to update these states across turns rather than regenerate them wholesale.

H DecL

MinTL tracks dialogue-state changes with minimal Levenshtein belief spans, then uses the updated state and knowledge-base results to generate responses. Its encoder-decoder framework jointly trains state-update and response generation components with pretrained sequence-to-sequence models.

  • Levenshtein belief spans: Levenshtein belief spans encode minimal slot-level edits—insertions, deletions, or substitutions—to update the previous dialogue state.NULL denotes deletion, while domain and slot information specifies the affected state entry.
  • Levenshtein belief spans: At inference, MinTL generates Levt and deterministically applies it to Bt−1, updating new values and deleting slots when NULL is produced.The described editing process inserts values such as people=10 and deletes hotel-area when the corresponding NULL operation is generated.
  • MinTL framework: The encoder receives the previous dialogue state and context, while a Lev decoder attends to the resulting hidden states to generate the current state edits.The context and state subsequences are concatenated with special segment tokens before encoding.
  • MinTL framework: The updated dialogue state queries external knowledge and booking APIs, whose categorized result selects a learnable KB-state embedding for response decoding.The selected embedding initializes the response decoder, which generates a delexicalized response.
  • Training and backbones: MinTL jointly optimizes Lev generation and response generation losses and can initialize its encoder and decoders with pretrained BART or T5 weights.Response generation is conditioned on the dialogue context, previous state, and knowledge-base result rather than the updated state directly.

4 Experiments

The experiments evaluate MinTL on the MultiWOZ multi-domain benchmark using pretrained sequence-to-sequence backbones and standard response-quality metrics.

  • MultiWOZ contains 8438, 1000, and 1000 training, validation, and testing dialogues across seven domains.
  • The evaluation tests dialogue state tracking on MultiWOZ 2.0 and 2.1, while end-to-end models are evaluated on MultiWOZ 2.0.
  • MinTL is instantiated with T5-small, T5-base, and BART-large pretrained sequence-to-sequence models.
  • End-to-end response quality is measured using Inform rate, Success rate, BLEU, and their combined score.The combined score is defined as (Inform + Success)×0.5 + BLEU.

4.4 Baselines

The baselines include oracle-state and pretrained language-model systems, recurrent and modular dialogue architectures, and alternative belief-state frameworks evaluated in low-resource and ablation settings.

  • End-to-end Modeling: End-to-end baselines include Oracle DST, HRED-TS, SFN + RL, MD-Sequicity, DAMD, DAMD + multi-action, and Sequicity + T5.
  • Low Resource Settings: The low-resource experiments use 5%, 10%, and 20% of the training data, corresponding to 400, 800, and 1600 dialogues.
  • Ablation Study: The ablation study compares MinTL variants using the original belief span and shared decoder parameters for Lev and response generation.
  • Dialogue State Tracking: DST comparisons include classification-based and generation-based systems, including MDBT, GLAD, SUMBT, TOD-BERT, TRADE, COMER, SOM-DST, DSTQA, and NADST.

4.5 Results

MinTL improves end-to-end response generation and low-resource performance, while Lev improves inference efficiency; its DST accuracy remains slightly below a classification-based state-of-the-art model on MultiWOZ 2.1.

  • End-to-end Modeling: MinTL achieves the best Inform rate, Success rate, and BLEU in end-to-end evaluation, improving the previous SOTA by around 10% in Success rate.With the same T5 backbone, MinTL achieves an around 11% higher Success rate than Sequicity.
  • Low Resource Settings: With only 20% training data, MinTL-based systems achieve competitive results compared to the full-data DAMD model.
  • Ablation Study: Replacing Lev with a standard belief span hurts overall performance, supporting Lev's effectiveness in the ablation study.
  • Dialogue State Tracking: The MinTL BART model achieves the highest joint goal accuracy among generation-based DST models on MultiWOZ 2.0 and 2.1.
  • Dialogue State Tracking: MinTL obtains a 1.62% lower joint goal accuracy than SST on MultiWOZ 2.1.The analysis attributes this comparison to classification-based models' advantage in predicting slot values from valid candidates.
  • Latency Analysis: MinTL is around 4 times faster than Sequicity, generating only 6 tokens per turn; it is also 15 times faster than TSCP and around 7 times faster than TRADE.

5 Conclusion

MinTL jointly learns dialogue state tracking and response generation with pretrained language models. Levenshtein belief spans reduce DST complexity and improve inference efficiency, while experiments on MultiWOZ report new state-of-the-art results.

  • MinTL is a simple, general framework that jointly learns dialogue state tracking and dialogue response generation.
  • Levenshtein belief spans reduce dialogue state tracking complexity and improve inference efficiency.
  • T5 and BART are incorporated as pretrained sequence-to-sequence backbones.
  • Experiments on MultiWOZ report new state-of-the-art results for dialogue state tracking and end-to-end response generation.

A KB States

The KB states are categorized using the number of matching entities and booking availability, with domain-specific thresholds for the train and other domains.

  • KB states are categorized by the number of matching entities and booking availability.
  • For the train domain, the thresholds are T1 = 1 and T2 = 3.
  • For other domains, the thresholds are T1 = 5 and T2 = 10.

B Hyper-parameters

The reported training hyper-parameters include context window size, learning rate, and learning-rate decay rate. Learning rate is decayed when validation performance does not improve, and models are trained on Nvidia V100 hardware.

  • The hyper-parameters include context window size w, learning rate lr, and learning-rate decay rate lr-decay.
  • The learning rate is decayed when validation-set performance does not improve.
  • All models are trained on Nvidia V100 hardware.

C Generated Examples

Tables 8–10 present MinTL-generated examples for T5-small, T5-base, and BART-large on MultiWOZ. The examples show dialogue states, delexicalized responses, and lexicalized responses.

  • The generated examples cover T5-small, T5-base, and BART-large backbones.
  • The examples are drawn from end-to-end response generation on the MultiWOZ test set.
  • Each table shows dialogue states, delexicalized responses, and lexicalized responses generated by the model.
Loading 2009.12005v2…