Improving Conversational Recommender Systems via Knowledge Graph based Semantic Fusion

Kun Zhou; Wayne Xin Zhao; Shuqing Bian; Yuanhang Zhou; Ji-Rong Wen; Jingsong Yu

Source-linked AI summary

Improving Conversational Recommender Systems via Knowledge Graph based Semantic Fusion

Kun Zhou, Wayne Xin Zhao, Shuqing Bian, Yuanhang Zhou, Ji-Rong Wen, Jingsong Yu

arXiv:2007.04032v1Jul 8, 2020cs.CLcs.AIcs.IR

TL;DR

CRS lacks sufficient conversational context and must bridge the semantic gap between natural-language expressions and item-level preferences. The paper fuses word- and entity-oriented knowledge graphs with Mutual Information Maximization, then uses the aligned representations in recommender and dialog components. Experiments demonstrate effectiveness on both recommendation and conversation tasks, while the study leaves broader external information, explainability, and historical profiles for future work.

Problem
CRS conversations provide limited context for understanding preferences, and natural-language utterances are semantically separate from item- or entity-level preferences.
Method
The approach uses ConceptNet and DBpedia, learns graph embeddings, aligns their word- and item-level spaces with Mutual Information Maximization, and builds KG-enhanced recommender and dialog components.
Results
The approach demonstrates effectiveness on both recommendation and conversation tasks.
Takeaways & Limitations
KG-based semantic fusion supports improved data representations for item recommendation and informative utterance generation within CRS.
Takeaways & Limitations
The study assumes no previous user-item interaction records and identifies user preference using conversation data only.

Abstract

from arXiv · show

Conversational recommender systems (CRS) aim to recommend high-quality items to users through interactive conversations. Although several efforts have been made for CRS, two major issues still remain to be solved. First, the conversation data itself lacks of sufficient contextual information for accurately understanding users' preference. Second, there is a semantic gap between natural language expression and item-level user preference. To address these issues, we incorporate both word-oriented and entity-oriented knowledge graphs (KG) to enhance the data representations in CRSs, and adopt Mutual Information Maximization to align the word-level and entity-level semantic spaces. Based on the aligned semantic representations, we further develop a KG-enhanced recommender component for making accurate recommendations, and a KG-enhanced dialog component that can generate informative keywords or entities in the response text. Extensive experiments have demonstrated the effectiveness of our approach in yielding better performance on both recommendation and conversation tasks.

1 INTRODUCTION

CRS must address sparse conversational context and the semantic gap between words and item-level preferences. The proposed approach fuses word- and item-oriented knowledge graphs to align representations and improve recommendation and response generation.

Motivation: CRS differs from traditional recommendation by combining interactive clarification and explicit natural-language feedback with recommender and dialog components.The recommender learns preferences from contextual utterances, while the dialog component clarifies intents and generates replies.
Motivation: Two unresolved problems are sparse conversational context and a semantic gap between natural-language utterances and item- or entity-level preferences.Understanding words such as “scary” and entities such as “Paranormal Activity (2007)” requires information not contained in the short utterances alone.
Approach: The approach incorporates ConceptNet for word relations and DBpedia for structured item facts, then uses graph neural networks to learn embeddings separately.ConceptNet captures relations such as synonyms, antonyms, and co-occurrence, while DBpedia supplies item attributes.
Approach: Mutual Information Maximization aligns the two knowledge-graph spaces by bringing word and item representations closer when they co-occur in conversations.This alignment connects contextual words with items and mentioned entities.
Contributions: The resulting system includes KG-enhanced recommender and dialog components for accurate recommendations and informative keywords or entities in responses.The design targets both item recommendation and utterance generation.
Results: Extensive experiments on a public CRS dataset demonstrate effectiveness on both recommendation and conversation tasks.

2 RELATED WORK

Related CRS research combines recommendation and dialogue to understand user needs, but prior methods often emphasize recommendation while using simpler conversation mechanisms. The paper responds by fusing word-level and entity-level knowledge graphs for both components.

CRS: CRS research integrates recommender and dialog components to recommend items through natural-language conversation.
Recommender systems: Traditional recommender systems rely on historical interactions, whereas CRS primarily uses conversational data rather than historical interaction data.Related recommendation methods also use item side information to address sparse interactions.
Conversation systems: Conversation-system research includes retrieval-based methods that select repository responses and generation-based methods that produce response text.
CRS: Earlier integrated CRS methods emphasize precise recommendation while implementing conversation with simple or heuristic solutions.A standard CRS dataset and hierarchical RNN generation model were later introduced.
This work: This work incorporates and fuses word-level and entity-level knowledge graphs to improve representations for recommendation and utterance generation.

3 PRELIMINARIES

A CRS analyzes multi-turn conversation history to infer user preferences, recommend items, or continue clarification. Its recommender and dialog components jointly produce the next recommendation set and response utterance.

CRS definition: CRS uses multi-turn conversation to analyze user preference and either recommend items or initiate another clarification round.The process ends when the task succeeds or the user leaves.
CRS definition: The recommender and dialog components should be integrated seamlessly, with successful recommendation as the final goal.
Formalization: A conversation history C is a list of utterances, with each utterance s_t representing the sentence at turn t.The formalization also distinguishes users u, items i, and vocabulary words w.
Formalization: At each turn, the recommender selects candidate items I_t, while the dialog component generates the next utterance s_t in response to prior utterances.I_t may be empty when clarification or chit-chat is needed instead of recommendation.
Formalization: For an n-turn conversation, the CRS outputs both the recommendation set I_n+1 and reply utterance s_n+1.

4 APPROACH

KGSF enhances conversational recommendation by encoding word- and item-oriented knowledge graphs, aligning their semantic spaces with Mutual Information Maximization, and using the fused representations for recommendation and response generation.

4.1 Encoding External Knowledge Graphs: KGSF encodes word and item semantic units with separate knowledge graphs to improve representations for conversational recommendation.It uses ConceptNet for word relations and DBpedia for item and entity facts.
4.1 Encoding External Knowledge Graphs: Graph neural networks learn node embeddings from ConceptNet and DBpedia, using GCN for word relations and R-GCN for relation-aware item representations.The item-oriented encoder retains relation semantics through relation-specific transformations.
4.2 KG Fusion via Mutual Information Maximization: Mutual Information Maximization aligns word and entity representations by treating co-occurring word–entity pairs as positive samples and random pairs as negative samples.A transformation matrix maps the two embedding spaces before the mutual-information objective is optimized.
4.2 KG Fusion via Mutual Information Maximization: A conversation-level super token replaces exhaustive word–entity pairing, using self-attention to summarize contextual words and improve efficiency.The attention weights reflect the importance of individual words in the conversation.
4.4 KG-enhanced Response Generation Module: The fused knowledge-graph representations support both item recommendation and KG-enhanced utterance generation.The response module incorporates KG-enhanced context-word and item representations through KG-based attention layers.
4.3 KG-enhanced Recommender Module: The recommender combines self-attended word and item vectors through a gate to form user preference, then ranks items using their learned embeddings.The model infers preference from conversation data without previous user–item interaction records.

5 EXPERIMENT

The experiments establish the evaluation setting before presenting results and analysis.

The experiments begin by defining the experimental setup.
The paper then reports the experimental results.
The section concludes with analysis of those results.

5.1 Experiment Setup

The model is evaluated on REDIAL across recommendation and conversation tasks, using ranking, diversity, and human-evaluation measures under standard and cold-start settings.

Dataset: REDIAL contains 10,006 conversations, 182,150 utterances, and 51,699 movies, split into training, validation, and test sets at an 8:1:1 ratio.The dialogs were generated by Amazon Mechanical Turk workers in seeker–recommender pairs.
Evaluation Tasks: The evaluation covers recommendation and conversation, reflecting the two major tasks in conversational recommendation systems.
Baselines: Recommendation baselines include Popularity and TextCNN, while Transformer, REDIAL, and KBRD provide conversation-oriented comparisons.KGSF is the proposed model evaluated against these baselines.
Evaluation Metrics: Recommendation performance is measured with Recall@k for k = 1, 10, 50, including a cold-start setting without mentioned items in the context.The cold-start experiment examines whether fusion helps learn preference from word-based utterances.
Evaluation Metrics: Conversation quality is evaluated automatically with Distinct n-gram metrics for n = 2, 3, 4 and additionally through human evaluation.
Implementation: The implementation uses PyTorch, graph neural networks with one layer for each KG, Adam optimization, and an MIM-loss weight λ of 0.025 during fine-tuning.The embedding dimensionalities are 300 for the conversation module and 128 for the recommender module.

5.2 Evaluation on Recommendation Task

KGSF improves recommendation by combining word-oriented and item-oriented knowledge graphs, with the largest benefit appearing in the cold-start setting. MIM further improves training efficiency and stability.

All Data Setting: KGSF outperforms recommendation baselines in the all-data setting by combining word-oriented and entity-oriented KGs through semantic fusion.KBRD improves over ReDial by incorporating external KG information, while KGSF adds word-level KG information and fuses both representations.
Cold Start Setting: KGSF achieves the best recommendation performance in the cold-start setting, where no items or attributes appear in the context.Its aligned word- and item-level semantic spaces help infer user preferences from utterance text.
Ablation Analysis: Both MIM and DBpedia contribute to the final recommendation performance in ablation analysis.Removing either the MIM loss or the item-oriented DBpedia KG reduces performance.
Effect of MIM: MIM reaches good recommendation performance with fewer iterations than the variant without MIM and improves training stability.Figure 2 compares Recall@10 and Recall@50 on the test set for KGFS with and without the MIM loss.

5.3 Evaluation on Conversation Task

KGSF performs better than the baselines on conversation evaluation, producing more fluent and informative responses. Ablations show that its KG-enhanced decoder components and MIM loss each contribute to response quality.

Automatic Evaluation: KGSF is consistently better than all baselines across automatic conversation evaluation metrics.The model uses KG information in a knowledge-enhanced Transformer decoder and a copy mechanism to improve generated-text informativeness.
Human Evaluation: KGSF is consistently better than all baselines with a large margin in human evaluations of fluency and informativeness.The authors attribute this to effective use of contextual information in a KG-enhanced Transformer decoder.
Ablation Study: Removing KG-based attention, the copy mechanism, or MIM reduces conversation performance, with KG-based attention producing the largest decrease.KG-based attention injects fused KG information into the decoder through multi-head attention.
Ablation Study: MIM improves generated-response quality in addition to improving recommendation performance.This result supports its usefulness for KG-based semantic fusion across both CRS tasks.

5.4 Qualitative Analysis

A sampled conversation illustrates how KGSF uses KG-enhanced word attention to identify user preferences and generate related movie entities. The example shows recommendation success after an initially unsuccessful round.

Recommendation Attention: KGSF assigns greater attention to “fantasy,” enabling the recommender to return “Pan’s Labyrinth” for a fantasy-movie request.The attention weights are computed by self-attention over KG-enhanced word embeddings.
Response Generation: The dialog component generates both the recommended movie “Pan’s Labyrinth” and the related movie “Stardust” in its response.The example illustrates that the response includes additional movie information beyond the recommender’s direct candidate.
Response Generation: Words receiving low recommender attention can receive higher dialog attention when they help generate a more informative reply.The two components therefore use attention differently within the sampled conversation.

6 CONCLUSION AND FUTURE WORK

The paper concludes that KG-based semantic fusion improves CRS recommendation and conversation performance by enhancing word and item representations and aligning their semantic spaces. Future work will add further external information and pursue more persuasive, explainable utterances and user profiles.

Conclusion: KG-based semantic fusion enhances word and item representations and aligns the semantic spaces used by CRS components.The approach uses two external knowledge graphs and Mutual Information Maximization.
Conclusion: The KG-enhanced recommendation and dialog components yielded better performance than several competitive baselines.The reported improvements cover recommendation and conversation tasks.
Future Work: Future work will incorporate additional external information, such as user demographics, to improve CRS performance.
Future Work: Future work will investigate more persuasive and explainable recommendation utterances and integration of historical interactions with pre-learned user profiles.