Encoding syntactic representations with a neural network for sentiment collocation extraction

Sentiment collocation refers to the collocation of a target word and a polarity word. Sentiment collocation extraction aims to extract the targets and their modifying polarity words by analyzing the relationships between them. This can be regarded as a basic sentiment analysis task and is relevant in many practical applications. Previous studies relied mainly on the syntactic path, which is used to connect the target word and the polarity word. To deeply exploit the semantic information of the syntactic path, we propose two types of syntactic representation, namely, relation embedding and subtree embedding, to capture the latent semantic features. Relation embedding is used to represent the latent semantics between targets and their corresponding polarity words, and subtree embedding is used to explore the rich syntactic information for each word on the path. To combine the two types of syntactic representations, a neural network is constructed. We use a recursive neural network (RNN) to model the subtree embeddings, and then the subtree embedding and the word embedding are combined as the enhanced word representation for each word in the syntactic path. Finally, a convolutional neural network (CNN) is adopted to integrate the two types of syntactic representations to extract the sentiment collocations from reviews. Our experiments were conducted on six types of reviews, which included product domains (such as cameras and phones) and service domains (such as hotels and restaurants). The experimental results show that our proposed method can accurately capture the latent semantic features hidden behind the syntactic paths that neither the common feature-based methods nor the syntactic-path-based method can handle, and, further, that it significantly outperforms numerous baselines and previous methods.