PaperlayerView exact revision on arXiv

Source-linked AI summary

RevCore: Review-augmented Conversational Recommendation

Yu Lu, Junwei Bao, Yan Song, Zichen Ma, Shuguang Cui, Youzheng Wu, Xiaodong He

arXiv:2106.00957v1cs.CL

TL;DR

Conversational recommendation can lack sufficient item information in short dialogues, limiting preference capture and informative responses. RevCore incorporates sentiment-matched reviews into recommendation and response generation, and experiments report improvements in both recommendation and dialogue responding.

  • Problem

    Short dialogue histories can lack item information for capturing preferences, and conversational recommenders can struggle to provide informative item-related descriptions.

    Source: §1 · p. 1

  • Method

    RevCore retrieves sentiment-matched reviews and uses review-enriched item information for recommendations and review-attentive response generation.

    Source: §1 · p. 2, §6 · p. 9

  • Results

    RevCore achieves consistent and significant improvements over baselines in recommendation and dialogue responding.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The authors report that RevCore generates informative responses without losing fluency and coherence.

    Source: §6 · p. 9

Abstract

from arXiv · show

Existing conversational recommendation (CR) systems usually suffer from insufficient item information when conducted on short dialogue history and unfamiliar items. Incorporating external information (e.g., reviews) is a potential solution to alleviate this problem. Given that reviews often provide a rich and detailed user experience on different interests, they are potential ideal resources for providing high-quality recommendations within an informative conversation. In this paper, we design a novel end-to-end framework, namely, Review-augmented Conversational Recommender (RevCore), where reviews are seamlessly incorporated to enrich item information and assist in generating both coherent and informative responses. In detail, we extract sentiment-consistent reviews, perform review-enriched and entity-based recommendations for item suggestions, as well as use a review-attentive encoder-decoder for response generation. Experimental results demonstrate the superiority of our approach in yielding better performance on both recommendation and conversation responding.

1 Introduction

Conversational recommendation is constrained by short dialogue histories and limited item-related descriptions, while knowledge-graph approaches have construction and integration drawbacks. RevCore uses sentiment-matched reviews to enrich recommendations and responses, and the authors report gains on both recommendation accuracy and conversation quality.

1 Introduction

  • Existing conversational recommenders face short dialogues with insufficient preference information and difficulty generating informative item-related descriptions.Knowledge-graph methods improve CR to some extent but have construction-cost and knowledge-integration limitations.Source: §1 · p. 1
  • Reviews provide detailed comments on users’ interests, while unfamiliar items can leave systems giving uninformative replies and few explanatory details.The paper presents reviews as external information that may support higher-quality recommendations and descriptive responses.Source: §1 · pp. 1–2
  • RevCore retrieves reviews matching the sentiment of mentioned items, incorporates them into dialogue history, and uses them to inform recommendation and response generation.A review-attentive decoder introduces selected-review information into responses.Source: §1 · p. 2
  • RevCore is reported to improve both recommendation accuracy and conversation quality over baselines.The authors also report informative responses without losing fluency and coherence.Source: §1 · p. 2

2 The Proposed Framework

RevCore retrieves sentiment-matched reviews and integrates them with dialogue context to support recommendation and response generation. Its three components retrieve reviews, recommend items from review-enriched entities, and generate review-informed responses.

2 The Proposed Framework

  • The framework takes dialogue context and retrieved reviews as inputs for both item recommendation and response generation.The review retrieval module selects reviews from a database; the recommender outputs candidate items, and the dialogue component outputs the next utterance.Source: §2 · p. 2

2.1 Review Retrieval

  • Sentiment-aware retrieval selects relevant reviews for mentioned items, while review selection can trade sentence fluency for greater word or phrase variability.The module filters for sentiment-consistent reviews and retains one sentence per mentioned movie; word- or phrase-wise selection offers more variability despite reduced fluency.Source: §2.1 · p. 3

2.2 Review-augmented Recommendation

  • Retrieved reviews enrich dialogue entities used to form a user representation and predict recommendations.The recommender extracts entities from context and reviews, aggregates entity embeddings into a user embedding, then predicts over candidate items.Source: §2.2 · pp. 3–4

2.3 Review-augmented Response Generation

  • The response generator separately encodes context and reviews, then fuses review information during decoding and uses copy probabilities from knowledge-graph entities and reviews.A copy mechanism adds vocabulary bias to increase informativeness; context encoding is intended to maintain dialogue consistency.Source: §2.3 · p. 4

3 Experiment Settings

Experiments use REDIAL conversations and IMDb reviews, with review-aware models compared against prior systems on recommendation and conversation tasks. Evaluation combines recommendation recall, automatic language metrics, and human ratings of response quality.

3.1 Dataset

  • REDIAL contains 10,021 conversations about 64,362 movies, and the review database uses the 30 highest-helpfulness IMDb reviews per movie.The conversations are seeker–recommender pairs; the dataset is split into training, validation, and test sets at an 8:1:12 ratio.Source: §3.1 · p. 5

3.2 Implementation Details

  • Implementation settings include context and response maximum lengths of 256 and 30 tokens, respectively, and 20 incorporated review tokens.Review sentences are selected by sentiment and sentence-wise; the passage notes other settings are consistent with Zhou et al. (2020) for fair comparison.Source: §3.2 · p. 5

3.3 Baselines

  • The evaluation compares RevCore with competitive prior systems on the REDIAL dataset.Baselines include Trans, Redial, KBRD, and KGSF.Source: §3.3 · p. 5

3.4 Evaluation Metrics

  • Recommendation is evaluated with Recall@1, Recall@10, and Recall@50; conversation uses perplexity, distinct 3- and 4-grams, and human ratings.Annotators score coherence, fluency, and informativeness from 0 to 1; the authors use human evaluation rather than BLEU because the dialogue component targets successful recommendation rather than ground-truth response imitation.Source: §3.4 · pp. 5–6

4 Results and Analysis

RevCore outperforms competing baselines on recommendation and conversation tasks, with review retrieval and response-generation components contributing to its results. The analyses also examine review length, retrieval strategy, and how reviews shape recommendations and generated responses.

4.1 Evaluation on Recommendation Task

  • RevCore achieves 5.9% R@1, 24.0% R@10, and 41.3% R@50, outperforming competitive baselines on REDIAL.The paper reports state-of-the-art recommendation performance; RevCore (−KG) also achieves competitive results and outperforms KGSF.Source: §4.1 · p. 6

4.2 Evaluation on Conversation Task

  • RevCore improves all Dist-n scores over KGSF by +0.14 for Dist-2, +0.11 for Dist-3, and +0.08 for Dist-4, with comparable PPL.RevCore (−KG) has higher Dist-n scores than KGSF but slightly lower scores than RevCore (+KG).Source: §4.2 · p. 6, Table 2 · p. 6
  • Human ratings place RevCore, with or without KG, above all baselines, especially for informativeness.Three annotators scored responses from 100 randomly selected multi-turn test dialogues.Source: §4.2 · p. 7, Table 3 · p. 7
  • Removing review copying, review attention, or the sentiment-aware review encoder reduces performance on generating diverse utterances.Copying and review attention appear more important for conversation diversity; sharing the review and context encoders remains an alternative, with only a slight increment from separated encoders.Source: §4.2 · p. 7, Table 4 · p. 7

4.3 Case Study

  • Sentiment-aware retrieval filters negative reviews when the user expresses a positive attitude, while incorporated reviews add entities and explanatory details to generated responses.The case study shows varied expressions and movie-related details in generated responses.Source: §4.3 · p. 7, Figure 3 · p. 8

4.4 Discussion

  • Review sentences of 20 words perform best across the tested metrics; longer reviews provide no further benefit.The authors suggest that excess external text may suppress the original text.Source: §4.4 · p. 7, Table 5 · p. 7
  • Randomly matched reviews lower R@k and Dist-n; sentiment-value ranking beats helpful-score ranking, and sentence-wise reviews have lower PPL than word-wise reviews.The authors also report that a food-review corpus yields the lowest R@k, despite not-bad Dist-n scores, and attribute negative recommendation effects to unrelated entities.Source: §4.4 · p. 8, Table 6 · p. 8

5 Related Work

The related work traces conversational recommendation from traditional recommender systems and dialogue-based interaction toward approaches that add external information. RevCore offers item-relevant reviews as an alternative source for enriching conversational recommendation.

5 Related Work

  • Traditional recommender systems estimate preferences from historical user-item interactions, using methods such as neural factorization machines and deep interest networks.Graph-based methods also model relations among users, items, and attributes.Source: §5 · p. 8
  • Conversational recommendation dynamically obtains user preferences through dialogue, with prior work exploring external knowledge, social networks, bandit strategies, multi-goal planning, explainability, and rapport-building.The cited work also includes a standard human-to-human, multi-turn movie-recommendation dataset.Source: §5 · pp. 8–9
  • RevCore differs from prior work by incorporating item-relevant reviews to learn user representations from review-enriched dialogue context for recommendation and response generation.Source: §5 · p. 9

6 Conclusion

RevCore combines review retrieval, recommendation, and review-informed response generation. The authors report improvements over baselines on recommendation and dialogue responding, with informative responses that retain fluency and coherence.

6 Conclusion

  • RevCore combines sentiment-aware review retrieval, recommendation from a review-enriched user profile, review encoding, and a review-attentive decoder for response generation.Source: §6 · p. 9
  • The authors report consistent and significant improvements over baselines in both recommendation and dialogue responding.Source: §6 · p. 9
  • RevCore generates informative responses without losing fluency and coherence.Source: §6 · p. 9

A Statistics for Conversation Dataset and Reviews

The REDIAL conversation statistics describe review coverage, vocabulary growth, and the movie sample used to link reviews. The study also tracks disliked and unseen recommendations.

A Statistics for Conversation Dataset and Reviews

B Experiment Details

The experiment details largely follow KGSF’s hyperparameters and training approach. The appendix also identifies the dataset statistics and comparison tables for model results and computational costs.

B Experiment Details

  • Most hyperparameters are kept consistent with KGSF, and the authors do not search additional combinations beyond their main idea.The recommender and dialogue embedding dimensions are 128 and 300, respectively; batch size is 32, learning rate is 0.001, and training runs for 30 epochs.Source: B Experiment Details · pp. 12–13
  • Table 7 lists REDIAL statistics for incorporated reviews, with “Cnd” and “Mnt” denoting candidate and mentioned movies.Source: Table 7 · p. 13
  • Table 8 compares three models by parameter count and training and inference time.Source: Table 8 · p. 13
  • Training proceeds by pre-training a sentiment predictor, training the recommender, and then training the dialogue component.The dialogue component is trained with other modules’ parameters frozen; the described training approach also pre-trains KG entity representations.Source: B Experiment Details · p. 13
  • Tables 9 and 10 report validation and test results for RevCore with and without a knowledge graph, including recall-rate and distance-score means and deviations.Source: Table 9 · p. 13, Table 10 · p. 13

C Model Size and Running Speed

The section reports model size and running speed for KGSF, RevCore (+KG), and RevCore (−KG), along with their shared experimental setup.

D Results on the Validation Set

The section provides validation results for RevCore with and without KG on REDIAL and points readers to the validation and test results in Table 9.

E Mean and Standard Deviation

The major experiment was repeated four times to assess RevCore's performance variability across metrics; the paper reports mean results for recommendation accuracy and conversation quality.

Related papers

Loading 2106.00957v1…