[Paper Review] Generate Natural Language Explanations for Recommendation

Hanxiong Chen, Xu Chen+2 more|arXiv (Cornell University)|Jan 9, 2021

Topic Modeling37 references37 citations

TL;DR

The paper proposes a hierarchical sequence-to-sequence model with an auto-denoising mechanism to generate free-text, personalized explanations for recommendations, while maintaining or improving prediction accuracy.

ABSTRACT

Providing personalized explanations for recommendations can help users to understand the underlying insight of the recommendation results, which is helpful to the effectiveness, transparency, persuasiveness and trustworthiness of recommender systems. Current explainable recommendation models mostly generate textual explanations based on pre-defined sentence templates. However, the expressiveness power of template-based explanation sentences are limited to the pre-defined expressions, and manually defining the expressions require significant human efforts. Motivated by this problem, we propose to generate free-text natural language explanations for personalized recommendation. In particular, we propose a hierarchical sequence-to-sequence model (HSS) for personalized explanation generation. Different from conventional sentence generation in NLP research, a great challenge of explanation generation in e-commerce recommendation is that not all sentences in user reviews are of explanation purpose. To solve the problem, we further propose an auto-denoising mechanism based on topical item feature words for sentence generation. Experiments on various e-commerce product domains show that our approach can not only improve the recommendation accuracy, but also the explanation quality in terms of the offline measures and feature words coverage. This research is one of the initial steps to grant intelligent agents with the ability to explain itself based on natural language sentences.

Motivation & Objective

• Motivate the need for personalized textual explanations to enhance transparency, persuasiveness, and trust in recommender systems.
• Address limitations of template- and retrieval-based explanations by enabling free-text explanation generation.
• Develop a model that can denoise noisy review data and tailor explanations to individual users and items.
• Jointly learn rating prediction and explanation generation in a multi-task framework to leverage shared representations.

Proposed method

• Introduce a hierarchical GRU-based generator consisting of a context GRU and a sentence GRU to produce multiple explanation sentences.
• Incorporate a feature-aware attention mechanism over product feature words extracted from reviews to personalize explanations.
• Apply an auto-denoising mechanism by weighting sentence losses with a supervised factor based on the proportion of feature words in each sentence.
• Use a word-level GRU with user and item embeddings to generate explanations, and initialize with a context-derived state.
• Train a multi-task objective that combines rating regression loss with weighted sentence generation losses and regularization.
• Evaluate using traditional NLP metrics (BLEU, ROUGE) and a feature-words coverage measure to reflect explainability.

Experimental results

Research questions

• RQ1 Can the hierarchical generation framework produce personalized natural language explanations without templates or retrieval from existing text?
• RQ2 Does incorporating feature-word based attention and auto-denoising improve the quality of explanations and the usefulness of the explanations for users?
• RQ3 How does the joint training of rating prediction and explanation generation impact recommendation performance and explanation quality?

Key findings

• The model uses offline metrics BLEU, ROUGE, and feature coverage to assess explanation quality alongside rating prediction performance.
• The auto-denoising mechanism helps focus training on sentences with higher feature-word content, improving explanation relevance.
• Feature-aware attention enables personalization by emphasizing relevant product features for each user-item pair.
• Experiments on Electronics and Beauty datasets demonstrate the approach’s viability for generating explanations while leveraging review data.