CoRE-CoG: Conversational Recommendation of Entities using Constrained Generation

Harshvardhan Srivastava, Kanav Pruthi, Soumen Chakrabarti, MausamarXiv 2023 · DOI · arXiv

Abstract

CoRE-CoG is a novel framework for conversational recommendation designed to bridge the gap between entity retrieval and natural language generation. By combining a recommendation "trigger," type-aware filtering, and a unique bidirectional decoder named HopSkip, the system ensures that recommended entities are fluently integrated into dialogue. Experimental results on benchmark datasets show a significant jump in Entity F1 scores and retrieval metrics over state-of-the-art baselines.

Introduction: The Retrieval-Generation Mismatch

Building an effective Conversational Recommendation System (CRS) requires balancing three critical tasks:

  1. Timing: Determining when to recommend versus when to seek more information.
  2. Relevance: Selecting the optimal entity from a large-scale knowledge base (KB).
  3. Fluency: Seamlessly weaving the selected entity into a natural response.

A common failure mode in current systems is the "handshake" between retrieval and generation. A model might correctly identify a movie but fail to mention it in the generated text, or recommend a category that contradicts the user's intent. This disconnect undermines the user experience, making the system appear illogical.

Methodology

CoRE-CoG (Conversational Recommendation of Entities using Constrained Generation) addresses these issues by extending the MESE architecture with three specialized modules:

Recommendation Trigger and Type-Cognizant Retriever

The system first employs a sequence-to-binary classifier to decide if a recommendation is warranted based on dialogue history. If triggered, a Type-Cognizant Retriever predicts the entity's category (e.g., "Movie" vs. "Food") before searching the KB. This hard filtering step reduces search space, improves speed, and prevents jarring category errors.

HopSkip: The Constrained Decoder

The core innovation is HopSkip, a bidirectional autoregressive decoder. Traditional language models generate text left-to-right, which often leads to poor "lookahead" when a specific entity must be included. HopSkip reverses this logic: it starts with the target entity and expands the sentence outward in both directions.

The HopSkip decoder builds sentences bidirectionally around a target entity to ensure inclusion and fluency.

By centering the generation process on the entity, the model guarantees its inclusion while maintaining high linguistic quality.

Experimental Setup

The framework was evaluated on two major benchmarks:

  • ReDial: A large-scale movie recommendation dataset.
  • DuRecDial 2.0: A multi-domain dataset (movies, music, food, restaurants). The authors curated a high-quality Knowledge Base for this dataset to ensure robust entity linking.

Results and Discussion

CoRE-CoG consistently outperformed top-tier baselines, including MESE and KBRD, across both retrieval and generation metrics.

Performance comparison of CoRE-CoG against baselines on ReDial and DuRecDial 2.0 datasets.

Key Findings

  • Retrieval Accuracy: On DuRecDial 2.0, the model achieved a 4 percentage point gain in Recall@1 and a notable boost in Mean Reciprocal Rank (MRR).
  • Generation Consistency: The most significant improvement was in Entity F1, which tracks how accurately the system mentions the intended entities. CoRE-CoG saw gains of up to 10 F1 points, as the HopSkip decoder prevents the model from "forgetting" the entity during generation.
  • Human Preference: Blind human evaluations rated CoRE-CoG higher in both fluency and informativeness compared to the MESE baseline.

Conclusion and Future Work

The success of CoRE-CoG highlights the power of hard constraints and bidirectional decoding in task-oriented dialogue. By treating entity inclusion as a requirement rather than a soft goal, the system achieves superior reliability. Future work will address the "cold-start" problem in early dialogue turns and explore scaling the architecture to open-world knowledge bases like Wikipedia.

Summary updated

Abstract

End-to-end conversational recommendation systems (CRS) generate responses by leveraging both dialog history and a knowledge base (KB). A CRS mainly faces three key challenges: (1) at each turn, it must decide if recommending a KB entity is appropriate; if so, it must identify the most relevant KB entity to recommend; and finally, it must recommend the entity in a fluent utterance that is consistent with the conversation history. Recent CRSs do not pay sufficient attention to these desiderata, often generating unfluent responses or not recommending (relevant) entities at the right turn. We introduce a new CRS we call CoRE-CoG. CoRE-CoG addresses the limitations in prior systems by implementing (1) a recommendation trigger that decides if the system utterance should include an entity, (2) a type pruning module that improves the relevance of recommended entities, and (3) a novel constrained response generator to make recommendations while maintaining fluency. Together, these modules ensure simultaneous accurate recommendation decisions and fluent system utterances. Experiments with recent benchmarks show the superiority particularly on conditional generation sub-tasks with close to 10 F1 and 4 Recall@1 percent points gain over baselines.
BibTeX
@misc{srivastava2023core,
  author = {Harshvardhan Srivastava and Kanav Pruthi and Soumen Chakrabarti and Mausam},
  title = {{CoRE-CoG: Conversational Recommendation of Entities using Constrained Generation}},
  year = {2023},
  eprint = {2311.08511},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.08511},
}
Figure