Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
38 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
41 tokens/sec
o3 Pro
7 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Towards Conversational Recommendation over Multi-Type Dialogs (2005.03954v3)

Published 8 May 2020 in cs.CL and cs.AI

Abstract: We propose a new task of conversational recommendation over multi-type dialogs, where the bots can proactively and naturally lead a conversation from a non-recommendation dialog (e.g., QA) to a recommendation dialog, taking into account user's interests and feedback. To facilitate the study of this task, we create a human-to-human Chinese dialog dataset \emph{DuRecDial} (about 10k dialogs, 156k utterances), which contains multiple sequential dialogs for every pair of a recommendation seeker (user) and a recommender (bot). In each dialog, the recommender proactively leads a multi-type dialog to approach recommendation targets and then makes multiple recommendations with rich interaction behavior. This dataset allows us to systematically investigate different parts of the overall problem, e.g., how to naturally lead a dialog, how to interact with users for recommendation. Finally we establish baseline results on DuRecDial for future studies. Dataset and codes are publicly available at https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/Research/ACL2020-DuRecDial.

Conversational Recommendation in Multi-Type Dialogs

This paper presents the novel task of conversational recommendation within multi-type dialogs, setting the stage for new developments in conversational AI. Previous research on dialog-based recommendation systems has focused primarily on either task-oriented or non-task dialog models, with a predominant concentration on singular dialog types. This research seeks to advance beyond existing paradigms by integrating multiple dialog types within a single system, thus allowing conversational agents to transition seamlessly from non-recommendation dialogs, such as question-and-answer (QA) sessions or chit-chat, to recommendation-focused interactions.

The authors emphasize the proactive role of the conversational agent, or "bot," challenging the assumption that dialog participants inherently understand the conversational goal. For the proposed task, the bot is expected to guide conversations naturally towards recommendation goals based on user interests, utilizing a combination of long-term objectives, such as recommending a specific entity, and short-term dialog goals that promote fluid topic transitions.

To support this endeavor, the authors introduce DuRecDial, a human-to-human Chinese dialog dataset consisting of approximately 10,000 dialogs and 156,000 utterances. Each dialog demonstrates the dynamics between a recommendation seeker (user) and a recommender (bot), with the latter leading the conversation across different dialog types towards achieving a sequence of planned goals. Significant attention is paid to capturing the variability in recommendation interactions, accommodating user feedback and preferences through progressive dialog rounds.

Methodology and Framework

The paper introduces a Mixed-Goal Conversation Generation (MGCG) framework, characterized by two core components: goal planning and response generation, each leveraging advanced neural encoder-decoder architectures. The goal-planning module predicts whether current dialog goals are complete and sets subsequent goals, encompassing both dialog type and topic selection. Meanwhile, the response generation module is tasked with producing contextually appropriate, knowledge-informed responses that fulfill assigned dialog goals.

The framework is equipped with mechanisms for goal completion estimation, dialog type and topic prediction, and leveraging external knowledge graphs for informative response generation. The researchers employ robust CNN models for goal prediction and knowledge-grounded GRU-based architectures for response generation, experimenting with both retrieval-based and generation-based responding variants to evaluate system performance.

Evaluation and Results

Performance metrics include typical automatic evaluations such as BLEU, F1, perplexity, and distinctness measures, as well as hits@k and knowledge precision recall metrics, to assess the relevance, fluency, diversity, and effectiveness of knowledge usage in generated dialog responses. Human evaluations further scrutinize fluency, informativeness, proactivity, and dialog coherence, validating the significance of integrating goal-driven planning in recommendation dialogs.

The empirical results confirm the merit of goal-centric dialog systems; both retrieval-based and generation-based models achieved superior performance in automatic and human evaluations compared to traditional sequence-to-sequence models, showcasing enhanced dialog appropriateness and informativeness. However, challenges persist in effectively completing certain dialog types, such as QA and recommendation, suggesting avenues for methodological refinement.

Implications and Future Directions

This paper lays the groundwork for a deeper exploration of multi-type conversational systems, promoting AI models capable of richer, goal-oriented interactions across varied dialog contexts. The introduction of DuRecDial and associated findings call for further research into adaptive dialog modeling, nuanced user profiling, and real-time knowledge integration within conversational agents. Potential future developments might explore transfer learning for cross-domain dialog systems or the incorporation of more sophisticated user feedback loops to refine recommendation accuracy and user engagement dynamically. These strides are anticipated to advance the theoretical foundations and practical applications of conversational AI significantly.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (6)
  1. Zeming Liu (34 papers)
  2. Haifeng Wang (194 papers)
  3. Zheng-Yu Niu (10 papers)
  4. Hua Wu (191 papers)
  5. Wanxiang Che (152 papers)
  6. Ting Liu (329 papers)
Citations (180)
Youtube Logo Streamline Icon: https://streamlinehq.com