Description

Product Aspect Mining and Stance Detection using Deep Learning Approach

Abstract

What other people think about the product has always been a crucial piece of advice for majority of us during the process of smart purchase decision. Internet has become an excellent source of consumer opinions as people tend to voice out their views to other people in terms of product reviews. But, it is not only difficult for a customer to read all of the reviews and make an informed decision on whether to purchase that product or not, but also it is extremely difficult for the producer of the product to manage and keep track of customer opinions. Therefore, mining online reviews (opinion mining) has always been a hot research topic. This proposal mainly focuses on two related subtasks in Opinion Mining Framework, aspect extraction and aspect polarity identification which is an important challenge to determine strengths and weakness of any product. The LSTM based deep learning model will be used for the former task of extracting aspects of the product, whereas later task on each aspect is performed through stance detection based on a CNN based supervised learning model and the performance will be compared to the state-of-art solutions using standard evaluation metrics.

Keywords — aspect extraction, stance detection, opinion mining, product aspect mining, deep learning model.

Due to the rapid improvement in the technology of the internet billions of people participate to express their views about products and services. In the e-commerce industry buyers rely on other buyer reviews to come to a conclusion if he/she wants to buy that product [4]. But due to the large amount of such emotional reviews and posts on the Internet, it is impossible for users to digest such information manually. Therefore, an efficient and automatic mining methodology that will remove off unnecessary data and bring about the most discussed or high frequency general sentiments about a product or service or social issues is the need of the hour. Such a technique is nowadays termed as Opinion Mining based on aspects.

In general, opinion mining aims to extract a quintuple <e, a, s, h, t> [1] from texts, where ‘e’ is the entity or the target, ‘a’ is the aspect of the entity ‘e’, ‘h’ is the opinion holder, ‘t’ is the time when the opinion holder expresses her opinion on the entity ‘e’, and ‘s’ is the opinion which ‘h’ holds to the aspect ‘a’ of the entity ‘e’ at ‘t’. For example, process of opinion mining in the following review text “I bought a new iPhoneX today, the screen is great, but the voice quality is poor” and outputs two quintuples <iPhoneX, screen, great, I, today> and <iPhoneX, voice quality, poor, I, today>. However, not all the opinion mining tasks need to extract all the five elements in quintuple. For example, sentimental analysis is concern more about the sentiment polarity ‘s’ of the text, stance detection merely focusses to identify the opinion ‘s’ to the specific target ‘e’, and product aspect mining focuses on extracting the aspect ‘a’ and corresponding opinions from text. Thus the proposed Aspect depended opinion mining technique combined with stance detection provides an excellent analysis of both sentiment and strength along with the polarity of the user review [2] not like the traditional sentimental analysis, which gives an overall/general idea whether a user review is positive or negative or neutral[3].

Opinion mining from online customer reviews mainly consists of two tasks. First, aspects must be extracted. Then, opinion associated with respective aspect must be identified and then must be oriented. Finally, summary has to be produced from the sentence lists.

Aspect term extraction (ATE) is the task of selecting the aspects[5] of an entity/target upon which opinions have been expressed, thus it can be formalized as task of extracting triple <e,a,s>( e means target, a and s represent aspect and opinion respectively). For example, the review “iPhoneX has a great sound quality, I like it.”, will extract triple of the review as <iPhoneX, sound quality, great>.Aspect sentiment classification (ASC) is the task of figuring out the polarities expressed on these selected aspects in the text of opinion [5].According to the general opinion mining framework, stance detection can be defined as the task of extracting tuple <e, s> (e means target and s represents opinion) without considering other elements [10]. Itmainly focuses on detecting the user stance (favor, against) on a particular customer review. It is similar to sentiment analysis [2], but with big difference.In specific, sentiment analysis aims to identify the sentiment polarity (positive, negative) of the text while stance detection cares about the stance on the target.

The concept of mining aspects and corresponding opinions was first addressed by Hu and Liu (2004) [5] using information extraction techniques and based on aspect frequency. They introduced the distinction between implicit and explicit aspects.Majority of the previous works in ATE have either of which ways have used conditional random fields (CRFs) [6, 7] or linguistic patterns [8, 9]. Both of these mentioned methods have their own limitations: linguistic patterns need to be crafted by hand, and they crucially depend on the grammatical accuracy of the sentences,CRF is a linear model, so it needs a large number of features to work well.These creates a room for scope and motivation of exploring and implementing an improved phrase-level entity/target opinion mining with enhanced aspect identification, opinion word extraction and aspect polarity identification.

To aid real world applications, our aim is to solve Aspect term extraction and Aspect sentiment classification simultaneously.In this proposal, we overcome existing limitations by using a Long-Short Term Memory (LSTM), a non-linear supervised classifier approach for aspect and opinion tagging. Further, Stance detection aims at recognizing the holistic subjective disposition (favor, against) that the consumer holds. Thus, we further propose use of stance detection using Convolution Neural Networks (CNN), for identifying aspect polarity in customer reviews.

Both stance detection and product aspect mining might benefit various downstream applications, such as they could benefit latent customers by providing smart purchase decision, manufactures by providing the measurement of customer satisfaction [11] to adjust their manufacturing process and sales strategy.