Research On Aspect Category Classification For Question-answering Style Reviews

Aspect category classification is one fundamental subtask in aspect-level sentiment analysis,which aims at identifying the aspect category of a given product review.This can help businesses learn about product aspects that consumers are concerned about to discover potential commercial value,and help consumers screen out product reviews related to their concerned aspects to make purchase decisions.With the rapid development of e-commerce business in our country,a new quesiton-answering(QA)style reviewing form in many popular e-commerce platforms,namely customer questions and answers.Although some researchers have devoted their efforts on aspect category classification for traditional product reviews,the research on aspect category classification for QA-style reviews is still in its infancy.Thus,we first draw up some aspect-related annotation guidelines for QA-style reviews and manually annotate to build up a high-quality corpus tailored for the new task,i.e.,aspect category classification for QA-style reviews.Then,we further conduct an in-depth study on this task.The main study of this paper includes the following three aspects:First,in QA-style reviews,the question and answer texts are both short and the number of words contained in the question or answer text is very limited.Besides,there are different aspect categories in different sentences in the answer text.Thus,we propose an aspect category classification method based on multi-attention representation.The core idea of this approach is as follows: we firstly segment the answer text into different sentences based on Sentence Segmentation Algorithm to make sure that each sentence contains only one aspect category,and then leverage the multi-attention representation layer to match the question text with each sentence inside the answer text and obtain the multiple attention representations of the question text to extend the features of the question text for classification.The empirical studies demonstrate that our proposed approach can make better use of the relevant information in the answer text and outperform other neural network methods,such as convolutional neural networks.Second,there are different aspect categories in different sentences inside the question or answer text,and the question text and the answer text are not well matched.For this reason,we propose a hierarchical matching attention network to address the aspect category classification task.Specifically,we firstly perform sentence segmentation in order to segment both the question and answer texts into sentences and then construct(sentence,sentence)units in each QA-style review.Then,we leverage an innovative QA matching attention layer to encode these(sentence,sentence)units in order to match the sentences inside question and answer texts.Finally,we leverage a self-matching attention layer to capture the different important degrees of different(sentence,sentence)units in each QA-style review.Empirical studies demonstrate that our hierarchical matching attention approach can further improve the performance of the aspect category classification task for QA-style reviews,which also shows that the attention mechanism is very effective for our task.Third,aspect categories in three domains are induced from aspect terms,so the extraction of aspect terms can assist the prediction of aspect categories.Hence,we propose a method based on joint learning,which leverages the aspect term extraction task to improve the performance of the aspect category classification task.Specifically,we first obtain a common representation of the question text for the two tasks in the shared Bidirectional Long Short-Term Memory(BiLSTM)layer.Then,for the aspect term extraction task,we adopt the Conditional Random Field(CRF)layer to perform sequence labeling on the question text and extract the aspect term.For the aspect category classification task,we leverage the attention layer to capture the matching information between the question and answer text and explore the potentail aspect information to extend the features of the question text.The experimental results demonstrate that our proposed approach based on joint learning helps to improve the classification performance of our task.