Abstract in English:
Review texts constitute a valuable source for making system-generated recommendations both more accurate and more transparent. Reviews typically contain statements providing argumentative support for a given item rating that can be exploited to explain the recommended items in a personalized manner. We propose a novel method called Aspect-based Transparent Memories (ATM) to model user preferences with respect to relevant aspects and compare them to item properties to predict ratings, and, by the same mechanism, explain why an item is recommended. The ATM architecture consists of two neural memories that can be viewed as arrays of slots for storing information about users and items. The first memory component encodes representations of sentences composed by the target user while the second holds an equivalent representation for the target item based on statements of other users. An offline evaluation was performed with three datasets, showing advantages over two baselines, the well-established Matrix Factorization technique and a recent competitive representative of neural attentional recommender techniques.