基於面向學習之商品評分預測與解釋文本生成模型

Abstract

本論文提出一個基於面向學習概念的模型，用來進行商品評分預測及對評分的解釋文本生成，稱為LARGE (Learning Aspects-representation for Rating and Generating Explanation)模型。在模型的編碼器中我們設計可學習多面向特徵空間的神經層，由商品評論文內容及使用者嵌入向量學習出對應的面向特徵向量，除了用以提供評分預測，並將商品的面向特徵向量融入每次的解碼狀態，引導生成的評分解釋文本能聚焦於商品所具有的面向。LARGE模型採多任務學習方式進行訓練，透過結合兩個不同目標任務的損失函數進行整體參數優化，且在評分預測的損失函數，加入權重調整策略，以降低推薦系統中評分資料分布不均對預測效能的影響。本論文採用亞馬遜資料集中三個不同商品類別的資料進行測試，實驗結果顯示，LARGE模型比相關研究所提出的代表性模型NRT，有效提升在評分預測及文本解釋生成的效能。此外，LARGE在解釋文本敘述中的類別型面向詞涵蓋率，比需輸入指定面向詞的NRT擴展模型有更高的涵蓋率。In this paper, we proposed a recommendation model, called LARGE (Learning Aspects-representation for Rating and Generating Explanation), to perform aspect-based representation learning for both rating prediction and explanation generation. In the encoder, we designed a neural layer to learn multi-aspect representations from item reviews and user embedding for the task of rating prediction. In the task of explanation generation, we fused the learned aspect-based representation of items into each decoding state in order to guide explanation generation focus on the specific aspects of the item. The LARGE model is trained by a multi-task learning approach, where the parameters are tuned by optimizing a linear combination of the loss on the two target tasks. In addition, to reduce the bias of model training due to data unbalance, a weight adjustment strategy is applied to the loss function of rating prediction. The experiments are performed on 3 categories selected from the Amazon review dataset. The result of the experiments shows that the LARGE model significantly outperforms NRT on both tasks. Furthermore, to compare with an extension model of NRT using a given aspect word as model input, the rating explanation generated by LARGE has higher coverage on aspect words.