MASER: Multi-Order Attention and Semantic-Enhanced Representation Model for Complex Text Recommendation

Abstract

In some recommendation platforms, the recommended items are composed of the complex text, and the target users are also described by the complex text. These texts are usually long, highly specialized, logically structured, and have significant differences, such as recommending technical demands of enterprises to technology researchers. Although some recommendation methods based on text representation can be used to solve this problem, such as Convolutional Neural Networks (CNNs) and Long Short Term Memory (LSTM), they may encounter challenges from different perspectives, e.g., path connectivity of representations, and the relationship between representations and recommended items. The complex text recommendation is an important problem that remains largely unsolved. In order to overcome the aforementioned challenges, by taking the technology commercialization as an example, which aims to recommend demands to researchers, we propose a novel complex text recommendation model called Multi-order Attention and Semantic Enhanced Representation (MASER). By integrating additional information into text vector representationsuch as structural relationship information for extended keywords, and semantic information for entity description texts the proposed model enhances complex text recommendation effectiveness significantly. Extensive experiments have been conducted on real datasets, confirming the advantages of the MASER model and the attention mechanism's effectiveness on complex text recommendation.