Source-linked AI summary
Hello, It's GPT-2 -- How Can I Help You? Towards the Use of Pretrained Language Models for Task-Oriented Dialogue Systems
Paweł Budzianowski, Ivan Vulić
TL;DR
Task-oriented dialogue systems need substantial domain-specific labeled data, motivating transfer from pretrained language models. The paper uses a text-only GPT-based model that folds dialogue, belief, and database states into one sequence, and evaluates it on MultiWOZ against a strong neural baseline. Automatic results are slightly lower, but human evaluations find no significant difference, while the authors emphasize portability and rapid adaptation.
Problem
Task-oriented dialogue models must learn many capabilities from scarce amounts of task-specific data, creating a need for transfer from large unlabeled corpora.
Method
The paper fine-tunes pretrained GPT models using a text-only sequence-to-sequence formulation that includes dialogue context, belief state, and database state.
Results
Automatic evaluation falls slightly below a strong task-specific neural baseline, while user-centered evaluation finds no significant difference between the models.
Takeaways & Limitations
The framework’s simple text encoding supports portability and quick adaptation to constrained domains and domain-specific vocabularies.
Abstract
from arXiv · showhide
Data scarcity is a long-standing and crucial challenge that hinders quick development of task-oriented dialogue systems across multiple domains: task-oriented dialogue models are expected to learn grammar, syntax, dialogue reasoning, decision making, and language generation from absurdly small amounts of task-specific data. In this paper, we demonstrate that recent progress in language modeling pre-training and transfer learning shows promise to overcome this problem. We propose a task-oriented dialogue model that operates solely on text input: it effectively bypasses explicit policy and language generation modules. Building on top of the TransferTransfo framework (Wolf et al., 2019) and generative model pre-training (Radford et al., 2019), we validate the approach on complex multi-domain task-oriented dialogues from the MultiWOZ dataset. Our automatic and human evaluations show that the proposed model is on par with a strong task-specific neural baseline. In the long run, our approach holds promise to mitigate the data scarcity problem, and to support the construction of more engaging and more eloquent task-oriented conversational agents.
1 Introduction
The paper addresses task-oriented dialogue data scarcity with a GPT-based model that uses text-only input instead of explicit policy and generation modules. On MultiWOZ, it approaches a strong baseline while offering portability and adaptability across domains.
- Motivation: Task-oriented systems require complex, costly data collection and annotation, while open-domain alternatives can produce unreliable responses or lack sufficient task coverage.Modular systems require independently trained components; retrieval systems provide control but can be predictable and limited by existing responses.
- Approach: The proposed GPT-based model consumes the full dialogue context as raw text, including belief state, database state, and previous turns.It bypasses explicit dialogue management and domain-specific natural language generation modules.
- Approach: The model follows TransferTransfo and pretrained GPT models to support task-oriented dialogue across domains.The paper combines pretrained generative modeling with components including tokenization, multi-task learning, and probabilistic sampling.
- Results: Automatic evaluation falls slightly below a strong task-specific neural baseline, while human evaluation finds no significant difference between the models.The proposed framework trades a small amount of automatic performance for portability and easier adaptation.
- Results: The GPT-based model learns domain-specific tokens, supporting adaptation to focused domains without the paper’s intricate modular design.The result suggests that a text-only formulation can provide broad portability across domains.
2 From Unsupervised Pretraining to Dialogue Modeling
The paper asks whether generative language-model pretraining can transfer to task-oriented dialogue, where manually labeled domain-specific data is scarce. It builds on language modeling, Transformer decoding, and TransferTransfo-style multi-task learning.
- Research question: Task-oriented dialogue modeling requires substantial domain-specific manually labeled data, motivating transfer from large unlabeled corpora.The paper frames this as whether generative pretraining can enable task-oriented dialogue modeling.
- Language-model pretraining: Standard language-model pretraining predicts the next word from preceding words by maximizing likelihood over word sequences.This provides the pretraining objective used before downstream adaptation.
- Language-model pretraining: Transfer learning with pretrained language models and Transformer decoders has produced progress across downstream tasks.The paper uses this transfer-learning setup as the basis for dialogue modeling.
- TransferTransfo: TransferTransfo combines the language-model objective with next-utterance classification during multi-task fine-tuning.This framework transfers a generative pretrained GPT model to dialogue.
- TransferTransfo: TransferTransfo significantly improves automatic and human dialogue evaluations on Persona-Chat.The cited result covers all automatic dialogue evaluation metrics and human-subject evaluation.
- TransferTransfo: Dialogue-state embeddings distinguish tokens belonging to the first speaker from tokens belonging to the second speaker.These embeddings are learned during fine-tuning to adapt GPT’s input representation to two-interlocutor dialogue.
3 Domain Transfer for (Task-Oriented) Dialogue Modeling
The paper combines pretrained generative models with text-only representations to transfer task-oriented dialogue across domains. Belief and database states become part of the generator’s textual input, while sampling strategies target response diversity and coherence.
- Framework: The framework joins dialogue context, belief state, and database state without standalone policy or generation modules.Its token-level embeddings identify which interlocutor contributed each part of the input.
- Domain transfer: The paper presents itself as the first combination of existing components aimed at task-oriented dialogue modeling with pretrained generative models.The components discussed include tokenization, simple text input, and generation strategies.
- Domain transfer: Subword tokenization and learned domain-specific token embeddings help pretrained models handle out-of-vocabulary and domain-specific vocabulary.The paper connects these capabilities to adapting off-the-shelf architectures without new dialogue submodules.
- Simple Text-Only Input: Simple text formulations let the model incorporate belief state and knowledge-base state directly into a sequence-to-sequence generator.New information can be added as another part of the text-only input.
- Generation quality: Nucleus sampling restricts generation to a probability-mass subset and is tested as an alternative to greedy sampling.The paper investigates whether this produces more varied and semantically richer responses without hurting task performance.
4 Fine-Tuning GPT on MultiWOZ
The model is evaluated on MultiWOZ using text representations of dialogue history, belief state, and database state. The setup compares text-only pretrained GPT variants with a neural baseline using an oracle belief state.
- Dataset: MultiWOZ contains 7 domains and 10,438 human-human dialogues for evaluating transfer to constrained, focused dialogue tasks.The dataset is described as substantially larger than previously available task-oriented dialogue datasets.
- Input representation: The task converts belief and database states into text and concatenates them with dialogue history as the model’s full context.Database entries are represented by domain and entity-count information, while belief state uses domain, slot, and value text.
- Input representation: Two learned token embeddings identify whether context tokens come from the user or the system.These embeddings inform the Transformer attention layers about the two conversation parties.
- Training details: Fine-tuning uses open-source GPT and GPT-2 checkpoints with selected training hyperparameters and two candidates per sequence.The implementation follows the paper’s pretrained-checkpoint setup.
5 Results and Analysis
The evaluation compares text-only GPT-based models with an oracle-belief-state neural baseline on MultiWOZ using automatic and human measures. GPT2-M improves task metrics with nucleus sampling, while human judgments find no substantial quality difference from the baseline.
- Automatic evaluation: The evaluation compares an oracle-belief-state neural response-generation baseline with text-only GPT-based models on dialogue-context-to-text generation.The tested pretrained models are GPT, GPT2, and GPT2-M.
- Automatic evaluation: Automatic evaluation measures task completion through Inform and Success rate, alongside fluency measured by BLEU.Inform checks whether an appropriate entity is provided, while Success rate checks whether requested attributes are answered.
- Automatic evaluation: With greedy sampling, the baseline leads on task-related metrics, whereas GPT2-M achieves the highest BLEU score.The authors note that greedy sampling may contribute to the performance gap by emphasizing pretrained-model properties.
- Automatic evaluation: Nucleus sampling improves GPT2 task scores on Inform and Success metrics but consistently lowers BLEU scores across models.The authors attribute the BLEU decrease to increased variability that may reduce domain-specific token generation.
- Qualitative analysis: Only around 50% of dialogues are successful for both the baseline and GPT-based models, with no clearly distinct success patterns between them.The authors suggest that a ranking model could potentially support ensembling the approaches.
- Human evaluation: Human rankings place original responses above all neural models, while comparisons between GPT-based models and the neural baseline are mixed and show no substantial quality differences.GPT is strongly preferred over the baseline, whereas the opposite pattern appears for GPT2.
6 Conclusion
The paper takes an initial step toward using large pretrained generative models for multi-domain task-oriented dialogue. Encoding the necessary information as simple text supports quick adaptation to constrained domains and domain-specific vocabularies.
- Conclusion: The paper presents a first step toward leveraging large pretrained generative models for multi-domain task-oriented dialogue.The authors aim to inform future work that improves and simplifies task-oriented conversational systems.
- Conclusion: Encoding the necessary information as simple text enables quick adaptation to constrained domains and domain-specific vocabularies.The conclusion frames this simplicity as a property of the fine-tuning procedure.