PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Proactive Dialogue Systems: Problems, Methods, and Prospects

Yang Deng, Wenqiang Lei, Wai Lam, Tat-Seng Chua

arXiv:2305.02750v2cs.CLcs.AI

TL;DR

Most dialogue systems overlook proactivity, the ability to take initiative and anticipate effects, even though it can support more complicated tasks. This survey reviews proactive dialogue problems and designs across dialogue types, along with data, evaluation, and future challenges. It presents proactivity as a practical problem for conversational AI that merits further research and exploration.

  • Problem

    Most dialogue systems overlook proactivity, the ability to take initiative and anticipate effects, although it can support tasks requiring strategic and motivational interaction.

    Source: §1 · p. 1

  • Method

    The survey systematically reviews proactive dialogue problems, designs, data resources, and evaluation protocols across open-domain, task-oriented, and information-seeking dialogue.

    Source: §1 · pp. 1–2

  • Results

    The survey organizes recent proactive dialogue studies across three dialogue types and summarizes their problems, designs, data resources, and evaluation protocols.

    Source: §1 · pp. 1–2

  • Takeaways & Limitations

    The paper concludes that developing proactive dialogue systems requires increased research and further exploration.

    Source: §6 · p. 7

    Robust and effective evaluation metrics remain a critical challenge, and only a few recent studies investigate proactivity in hybrid dialogues.

    Source: §5 · p. 7

Abstract

from arXiv · show

Proactive dialogue systems, related to a wide range of real-world conversational applications, equip the conversational agent with the capability of leading the conversation direction towards achieving pre-defined targets or fulfilling certain goals from the system side. It is empowered by advanced techniques to progress to more complicated tasks that require strategical and motivational interactions. In this survey, we provide a comprehensive overview of the prominent problems and advanced designs for conversational agent's proactivity in different types of dialogues. Furthermore, we discuss challenges that meet the real-world application needs but require a greater research focus in the future. We hope that this first survey of proactive dialogue systems can provide the community with a quick access and an overall picture to this practical problem, and stimulate more progresses on conversational AI to the next level.

1 Introduction

Conventional dialogue systems usually follow user-led conversations, while proactive systems take initiative and anticipate impacts. This survey reviews proactivity across dialogue types, covering problems, designs, data, evaluation, and future challenges.

1 Introduction

  • Proactivity is the ability to create or control conversation by taking initiative and anticipating effects on the system or users, rather than only responding passively.The survey connects proactivity with improved engagement and service efficiency, and with tasks requiring strategic and motivational interaction.Source: §1 · p. 1
  • Proactive open-domain systems can guide topic discussions or constructively redirect problematic utterances, while task-oriented systems can handle conflicting objectives or offer unrequested information.The survey also covers clarification questions and user preference elicitation for information-seeking systems.Source: §1 · pp. 1–2
  • The survey identifies hybrid dialogues, robust evaluation protocols, and the ethics of proactivity as challenges and prospects for future research.It notes potential harm to users as a concern in balancing benefits to human-AI interactions.Source: §1 · p. 2
  • The survey reviews proactive dialogue problems and designs across open-domain, task-oriented, and information-seeking systems, alongside available data resources and evaluation protocols.It describes itself as the first survey focused on proactive dialogue systems, to the authors’ knowledge.Source: §1 · pp. 1–2

2 Proactive Open-domain Dialogue Systems

Proactive open-domain dialogue systems lead conversations toward target topics or redirect problematic utterances in a prosocial manner. The survey reviews their task formulations, methods, datasets, and evaluation approaches.

2.1 Target-guided Dialogues

  • Target-guided dialogue systems lead conversation toward a designated target topic or a predefined topic thread instead of consistently following user-oriented topics.The agent begins from an arbitrary topic and is expected to maintain smooth transitions while achieving the target.Source: §2.1 · p. 2, Problem Definition · p. 3

Methods

  • Target-guided dialogue work divides into topic-shift detection, topic planning, and topic-aware response generation.Methods include detecting topic drift, planning keyword or knowledge-graph transitions, and generating responses that bridge current and target topics.Source: Methods · p. 3

Datasets and Evaluation Protocols

  • TGC labels targets on existing Persona-Chat conversations, whereas DuConv uses human-human conversations generated from designated targets and linked knowledge-graph entities.The target may be a keyword, knowledge entity, or conversational goal.Source: Datasets and Evaluation Protocols · p. 3, Problem Definition · p. 3
  • Target-guided systems are assessed at turn and dialogue levels, including target precision and recall, human ratings of proactivity and smoothness, success rate, and turns to target.Dialogue-level evaluation commonly uses user simulators because real-user experiments are costly and complex.Source: Datasets and Evaluation Protocols · p. 4

2 Proactive Open-domain Dialogue Systems

  • Passive responses to unsafe, unethical, or toxic user statements can raise safety concerns, motivating systems that detect problematic utterances and respond constructively according to social norms.The task is framed as classifying a safety label and then generating a response to mitigate the problematic utterance.Source: §2.2 · p. 4

2.2 Prosocial Dialogues

  • Pro-social dialogue systems are framed as detecting problematic utterances and respectfully guiding the conversation according to social norms.The survey describes this work as a response to systems that may passively agree with unsafe, unethical, or toxic statements.Source: §2.2 · p. 4
  • Pro-social methods cover safety detection, rule-of-thumb generation, and prosocial response generation, with fine-grained safety labels intended to avoid treating sensitive utterances as unsafe or toxic.Generated rules of thumb may not themselves make suitable user-engaged responses; response generation can instead condition on relevant rules and dialogue context.Source: §2.2 · p. 4
  • MIC and ProsocialDialog provide resources for evaluating proactive dialogue, while evaluation uses classification and text-generation metrics alongside human ratings or trained classifiers for prosocial attributes.The attributes include agreement, respect, and fairness.Source: §2.2 · pp. 4–5

3 Proactive Task-oriented Dialogue Systems

Proactive task-oriented dialogue systems address settings where agents must pursue system goals or add useful, unrequested information beyond simply completing user-requested tasks. The survey reviews non-collaborative and enriched task-oriented dialogues, their methods, datasets, and evaluation.

3 Proactive Task-oriented Dialogue Systems

  • Proactive task-oriented systems handle non-collaborative tasks with competing user and system goals or enrich conversations with useful, unrequested information.Examples include negotiation and persuasion, as well as supplementary knowledge or chitchat.Source: §3 · p. 5

3.1 Non-collaborative Dialogues

  • Non-collaborative dialogue systems seek consensus by generating responses and strategies from the dialogue history, prior strategies, and task background.The background may include item descriptions or user profiles, and strategies may range from dialogue acts to fine-grained labels.Source: §3.1 · p. 5
  • Existing non-collaborative methods address strategy learning, user personality modeling, and persuasive response generation, while strategy learning also requires rapport and trust-building skills.Approaches include strategy-graph networks, personality inference, and modular or end-to-end persuasive response generation.Source: §3.1 · p. 5
  • CraigslistBargain and PERSUASIONFORGOOD are human-human benchmarks; evaluation commonly uses strategy-prediction accuracy, F1, and ROC AUC, plus human judgments of persuasiveness and task success.PERSUASIONFORGOOD includes user profiles and manual annotations of persuasion strategies and dialogue acts.Source: §3.1 · pp. 5–6

3.2 Enriched Task-oriented Dialogues

  • Enriched task-oriented dialogues require responses to be functionally accurate and socially engaging, with datasets and methods adding chitchat or knowledge-grounded chitchat.Combiner is a pipeline-based method intended to reduce interference between dialogue state tracking and knowledge-enriched response generation.Source: §3.2 · p. 6

4 Proactive Conversational Information Seeking Systems

Proactive conversational information-seeking systems initiate subdialogues to clarify queries or elicit preferences, addressing limits of passively responding to user queries. The survey reviews clarification-question pipelines and preference-elicitation methods, while noting a need for human-human benchmarks and systems that combine contextual understanding with proactive elicitation.

4 Proactive Conversational Information Seeking Systems

  • Proactive information-seeking systems initiate subdialogues to clarify ambiguity or elicit preferences, helping address complicated information needs that passive query response may not fulfill.Applications include conversational search, question answering, and recommendation.Source: §4 · p. 6

4.1 Asking Clarification Questions

  • Clarification-question systems predict whether a query is ambiguous, then select or generate a clarifying question; later systems combine need prediction, question generation, and question answering.NeuQS reranks retrieved questions, QCM generates questions using reinforcement learning, and UniPCQA uses a unified sequence-to-sequence formulation for three tasks.Source: §4.1 · p. 6

4.2 User Preference Elicitation

  • The “System Ask, User Respond” paradigm elicits recommendation preferences by predicting which item attribute to ask about next.PMMN incorporates user embeddings for turn-level next-question prediction; later work treats multi-turn asking as a reinforcement-learning decision process.Source: §4.2 · pp. 6–7
  • Question-driven recommendation evaluation measures next-attribute prediction and item ranking at the turn level, and recommendation success and average session turns at the dialogue level.The described measures include HR@k,t, MRR@k,t, MAP@k,t, NDCG@k,t, SR@t, and AT.Source: §4.2 · p. 7
  • Current preference-elicitation studies mainly use synthetic conversations from product reviews or purchase logs, leaving human-human benchmarks in demand for more robust and reliable evaluation.The survey also identifies combining contextual natural-language understanding and generation with proactive preference elicitation as a beneficial direction.Source: §4.2 · p. 7

5 Challenges and Prospects

The survey identifies hybrid dialogues, proactivity evaluation, and ethics as challenges for real-world proactive dialogue systems. It calls for smoother transitions across dialogue types, robust multidisciplinary evaluation, and attention to factuality, morality, and privacy.

Proactivity in Hybrid Dialogues

  • Hybrid dialogue systems must handle multiple and changing conversational goals, with natural transitions and overall quality maintained without performance loss on particular dialogue types.The survey describes hybrid systems as the closest form of real-world applications.Source: §5 · p. 7

Evaluation Protocols for Proactivity

  • Proactivity evaluation remains challenging: human assessment is common and costly, while robust protocols must evaluate language and dimensions from fields including HCI, sociology, and psychology.User simulators are described as a relatively efficient, less labor-intensive way to evaluate proactive interactions; goal completion and satisfaction are identified as essential metrics.Source: §5 · p. 7

Ethics of Conversational Agent’s Proactivity

  • Proactive systems raise ethical concerns around factuality of added information, aggressiveness in non-collaborative dialogue, and misuse of personal information.The survey also identifies hallucination, toxic language, and social bias as relevant concerns.Source: §5 · p. 7

6 Conclusion

The conclusion presents a comprehensive review of conversational-agent proactivity across dialogue types, representative public data resources, and open challenges. It frames further research and exploration as necessary to develop proactive dialogue systems.

6 Conclusion

  • The paper reviews problems and designs for conversational-agent proactivity across different dialogue systems.Source: §6 · p. 7
  • The review summarizes publicly available data resources and highlights open challenges in the field.Source: §6 · p. 7
  • The authors state that further research and exploration are needed to develop proactive dialogue systems.Source: §6 · p. 7

Related papers

Loading 2305.02750v2…