SentiMem: Attentive Memory Networks for Sentiment Classification in User Review

Abstract

Sentiment analysis for textual contents has attracted lots of attentions. However, most existing models only utilize the target text to mine the deep relations from text representation features to sentiment values, ignoring users’ historicalally published texts, which also contain much valuable information. Correspondingly, in this paper we propose SentiMem, a new sentiment analysis framework that incorporates user’s historical texts to improve the accuracy of sentiment classification.

In SentiMem, to exploit users’ interests and preferences hidden in the texts, we adopt SenticNet to capture the concept-level semantics; as for users’ temperaments, we combine multiple sentiment lexicons with multi-head attention mechanism to extract users’ diverse characters. Then, two memory networks: Interests Memory Network and Temperaments Memory Network are used to store information about users’ interests and temperaments respectively. Interests memory is updated in a first-in-first-out way and read by an attention mechanism to match the users’ most recent interests with the target text. Temperaments memory is updated in a forgetting-and-strengthening manner to match the gradual shift of human’s characteristics. Additionally, we learn a global matrix to represent these common features among human’s temperaments, which is queried when classifying a user’s new posted text. Extensive experiments on two real-world datasets show that SentiMem can achieve significant improvement for accuracy over state-of-the-art methods.

This work was supported by the National Key R&D Program of China [2018YFB1004700]; the National Natural Science Foundation of China [61872238, 61972254]; the Shanghai Science and Technology Fund [17510740200], and the CCF-Huawei Database System Innovation Research Plan [CCF-Huawei DBIR2019002A].