PaperlayerView exact revision on arXiv

Source-linked AI summary

SOLOIST: Building Task Bots at Scale with Transfer Learning and Machine Teaching

Baolin Peng, Chunyuan Li, Jinchao Li, Shahin Shayandeh, Lars Liden, Jianfeng Gao

arXiv:2005.05298v4cs.CLcs.AIcs.LG

TL;DR

Building task bots is costly, and existing approaches require large amounts of task-specific labeled data that are rarely available for new tasks. SOLOIST pretrains a task-grounded dialog model on heterogeneous corpora and adapts it with a handful of teacher-generated dialogs. It achieves state-of-the-art benchmark performance and outperforms existing methods in few-shot settings.

  • Problem

    Building dialog systems at scale is costly, while existing methods require large amounts of task-specific labeled data that are rarely available for new real-world tasks.

    Source: §1 · p. 1

  • Method

    SOLOIST uses a Transformer-based autoregressive model pretrained for task-grounded response generation and adapts it to new tasks with teacher-generated dialogs.

    Source: §1 · p. 2

  • Results

    SOLOIST achieves state-of-the-art performance on two task-oriented dialog benchmarks and outperforms existing methods in few-shot fine-tuning settings.

    Source: §1 · p. 3, §5 · p. 12

  • Takeaways & Limitations

    The paper presents task-grounded pretraining followed by machine-teaching adaptation with a handful of task-specific examples as a way to build task bots at scale.

    Source: §5 · p. 12

    The approach assumes the database state is obtained by deterministic database lookup from the belief state.

    Source: §2.2 · p. 4

Abstract

from arXiv · show

We present a new method SOLOIST that uses transfer learning and machine teaching to build task bots at scale. We parameterize classical modular task-oriented dialog systems using a Transformer-based auto-regressive language model, which subsumes different dialog modules into a single neural model. We pre-train, on heterogeneous dialog corpora, a task-grounded response generation model, which can generate dialog responses grounded in user goals and real-world knowledge for task completion. The pre-trained model can be efficiently adapted to accomplish new tasks with a handful of task-specific dialogs via machine teaching, where training samples are generated by human teachers interacting with the system. Experiments show that (i) SOLOIST creates new state-of-the-art on well-studied task-oriented dialog benchmarks, including CamRest676 and MultiWOZ; (ii) in the few-shot fine-tuning settings, SOLOIST significantly outperforms existing methods, and (iii) the use of machine teaching substantially reduces the labeling cost of fine-tuning. The pre-trained models and codes are available at https://aka.ms/soloist.

1 Introduction

Building task bots at scale is costly and existing methods need large amounts of task-specific labeled data. SOLOIST uses a pretrained, task-grounded model and machine teaching to adapt bots, achieving strong benchmark and few-shot results.

1 Introduction

  • Building dialog systems remains label-intensive and time-consuming, requiring domain knowledge, coding skill, and expert experience; scaling to tens of thousands of bots can be prohibitively expensive.Existing data-driven and end-to-end methods also require large amounts of task-specific labeled data, rarely available for new real-world tasks.Source: §1 · pp. 1–2
  • SOLOIST parameterizes the sequential dialog pipeline with one Transformer-based autoregressive model pretrained on heterogeneous dialog corpora and adapted to new tasks using machine teaching.Human teachers generate task-specific training samples by interacting with the system.Source: Figure 1 · p. 2, §1 · p. 3
  • The approach transfers language understanding and generation learned from raw text alongside response grounding in user goals and real-world knowledge for task completion.The supplied empirical study reports these capabilities transfer from pretraining to a new task bot.Source: §1 · p. 3
  • SOLOIST achieves state-of-the-art results on two task-oriented dialog benchmarks, lifting the combined score by 10 points in automatic evaluation and success rate by 20 points in human evaluation.In few-shot settings, it achieves a reasonable success rate using less than 50 dialogs.Source: §1 · p. 3

2 SOLOIST

SOLOIST represents task-oriented dialog as one auto-regressive model that predicts belief states and generates responses grounded in dialog context and database state. It is pre-trained on task-oriented dialogs, then adapted to new tasks with labeled dialogs, including examples refined through machine teaching.

2.1 An Auto-Regressive Model for Dialog

  • A single auto-regressive model subsumes dialog modules by modeling the sequence produced as their inputs and outputs are concatenated.The sequence comprises dialog history, belief state, database state, and response.Source: §2.1 · p. 3

2.2 Task-Grounded Pre-Training

  • Each training turn is represented as dialog history, annotated belief state, retrieved database state, and delexicalized response, then tokenized for training.The example shows a restaurant query represented with a belief state, database match count, and response containing placeholders.Source: §2.1 · p. 3, §2.2 · pp. 4–5
  • The model predicts belief states from dialog history and generates responses conditioned on history, belief state, and database state.The database state is obtained deterministically from the belief state, splitting modeling into belief prediction and grounded response generation.Source: §2.2 · p. 4
  • The contrastive objective promotes matched positive sequences and drives down mismatched negatives, generated by replacing sequence items with different dataset items with probability 50%, and classifies sequences as matched or mismatched.Negative samples replace only the belief state, only the response, or both, with each type chosen with probability 1/3.Source: §2.2 · p. 4

2.3 Fine-Tuning and Machine Teaching

  • Machine teaching adapts the model using representative failed dialogs that authors revise by correcting beliefs or responses before fine-tuning.Authors select a dozen training samples from human-bot interaction logs; Figure 2 illustrates belief-state correction and response selection or insertion.Source: §2.3 · p. 5, Figure 2 · p. 5

3 Experiments

SOLOIST is evaluated on standard task-oriented dialog benchmarks, few-shot adaptation, and machine teaching. It performs strongly across these settings, while component-wise and human evaluations probe its task completion and interaction quality.

3.1 Experimental Setup

  • Evaluation covers CamRest676 and multi-domain MultiWOZ, four single-domain few-shot tasks, and machine teaching; Combined aggregates Inform, Success, and BLEU.Source: §3 · p. 6, §3.1 · p. 6, §3.3 · p. 7

3.2 End-to-End Evaluation

3.3 Few-Shot Evaluation

3.4 Machine Teaching Results

  • In Restaurant, teaching uses 61 versus 396 edits; five teacher-corrected dialogs improve Combined over no teaching and five fully annotated extra examples.Source: §3.4 · pp. 8–9, Table 8 · p. 9

3.5 Component-Wise Evaluation

3.6 Human Evaluation Results

  • Across 120 crowdworker sessions, SOLOIST outperforms other systems on all human metrics, with a Success margin 20 points above the DSTC8 Track 1 Winner.Source: §3.6 · pp. 10–11, Table 13 · p. 11

4 Related Work

Prior work spans task-oriented and social dialog systems, pretrained language models, and end-to-end dialog systems. SOLOIST extends pretrained approaches from individual dialog modules to an end-to-end task-bot pipeline.

4 Related Work

  • The paper focuses on end-to-end models for task-oriented dialog, within a broader field that also includes social chatbots and other specialized dialog systems.The passage names empathetic dialog, sentiment-analysis chatbots, commonsense-based systems, and systems using audio features.Source: §4 · p. 11
  • Pretrained dialog work commonly improves individual modules, while SOLOIST generalizes pretraining and fine-tuning to the entire end-to-end task-bot pipeline.Prior examples include TOD-BERT for dialog sub-tasks, ConveRT for intent classification and response selection, and Span-ConveRT for entity extraction.Source: §4 · p. 12
  • Earlier end-to-end systems were often domain-specific or required large amounts of in-domain labels for multi-domain tasks, whereas SOLOIST adapts in few-shot settings.The passage contrasts domain-specific end-to-end methods and label-intensive multi-domain methods with SOLOIST’s few-shot adaptation.Source: §4 · p. 12
  • SOLOIST differs from Ham et al. (2020) and Hosseini-Asl et al. (2020) through task-grounded pretraining, lower annotation needs, and end-to-end training.SOLOIST requires belief-state labels, while the compared methods require dialog-act labels for each turn; those methods also need heuristic rules for database search conditions.Source: §4 · p. 12

5 Conclusion

SOLOIST combines transfer learning and machine teaching to build task bots, using task-grounded pretraining to generate responses grounded in user goals and real-world knowledge. The reported experiments show state-of-the-art results on two task-oriented dialog benchmarks and strong performance in few-shot fine-tuning.

5 Conclusion

  • SOLOIST is a task-bot building method using transfer learning and machine teaching, with task-grounded pretraining for responses based on user goals and real-world knowledge.The paper reports new state of the art on two popular task-oriented dialog benchmarks and large-margin gains over existing methods in few-shot fine-tuning.Source: §5 · p. 12
  • The proposed recipe combines a single auto-regressive language model, heterogeneous-dialog pretraining, and fine-tuning on a handful of task-specific examples generated through machine teaching.The authors hope the recipe will inspire further exploration and improvement of this task-bot development paradigm.Source: §5 · p. 12

Related papers

Loading 2005.05298v4…