Comparative opinion contains contrasting views of products (e.g., which aspect of a product is better or worse). Most existing works for comparative opinion mining focus on single comparative sentences but have yet explored the benefits of additional comparative details in neighbouring sentences of a comparative sentence. Motivated by the needs to exploit these supporting details, this chapter proposes an approach to identify the link between a comparative sentence and its neighbouring sentences. As contextual similarity between comparative sentence and neighbouring sentence is crucial to determine their relatedness, contextual features of these two sentences are exploited to measure the similarity between them. Then, linear-chain conditional random field (CRF) is used to identify neighbouring sentence that is related to comparative sentence. Detection of supporting neighbouring sentence using linear-chain CRF with optimized contextual features (cosine similarity, Wordnet similarity, and comparative keywords) outperforms sentence similarity baseline by 4% in accuracy and 0.06 in F-score.
TopIntroduction
Growing number of online review websites such as Trip Advisor, Eopinions, Amazon, CNet etc. enables people to express their opinions about services and products. As a result, huge number of opinion-based texts were generated from these platforms. These resources are very useful for both service providers and end users as they can offer clues towards various aspects of the services or products. However, time and effort are needed for both parties to read and digest these opinionated resources, e.g., scanning through the sentences that carry opinions, filter and summarize them into simple conclusion like pros and cons, comparing between multiple features or attributes discussed, comparing between two (or more) different products or services and so forth.
To address such needs, research in opinion mining aims to make the opinions more explicit by summarizing the opinion using indicators like positive or negative. Combining both computational linguistics and text mining methods, opinion mining enables the detection of sentiment value expressed in textual form, which can be used as indicator of public responses on services, products or events (Hu & Liu, 2004). For example, in service industry like restaurant, feedback such as “The waiter is very friendly” can be classified as “positive” since the sentence contains positive sentiment word “friendly”. On the other hand, comment on product like “New IPhone is a lousy phone” reflects “negativity” with the presence of “lousy” as negative word (Pang & Lee, 2008).
Within the field of opinion mining, comparative opinion mining is an emerging sub area that gained attention in opinion mining research lately (Jindal & Liu, 2006; Jindal & Liu, 2006b; Khan et al, 2020; Xu et al, 2011, Yang & Ko, 2011; Younis et al, 2020). A recent review by Varathan et al, 2017 has also shown the increasing of research works in this area. Instead of dealing with opinions of sentence, which is either positive or negative opinionated, this sub area of opinion mining research looks into sentence that makes comparison between different entities with respect to a common feature for e.g., IPhone 8 has better camera compared to Samsung S8”. The sentence contains comparison relation, i.e., between IPhone 8 as first entity and Samsung S8 as second entity with regard to the common feature which is camera. In terms of opinion’s sentiment, “better” indicate IPhone 8 is more superior compared to Samsung S8 in camera feature. Looking at the example, it is obvious that this type of opinion is useful for decision making in terms of comparative analysis. Hence, in this work, we are motivated to explore research in the direction of comparative opinion mining.
In opinion mining, customer reviews often contain opinions that make comparison between two products or services with respect to some common features. This type of opinion usually provides contrasting views of different products. Sometimes, it could be making similar remarks, such as, “Mi Note is as good as Iphone.”. This type of review is particularly useful in situations where a consumer is making a decision to choose between two or more similar products.
From business point of view, comparative texts are important for companies to analyse and discover their competitors’ products as well as to learn their relative weaknesses and strengths of their products. Conventionally, such information is available through analyst reports and expert reviews, which mostly contain opinions based on personal experiences of the expert. Also, product press release could be biased as it is written by the product’s company.
In comparative opinion, sentiment analysis is normally based on the thought that relates some common features of two or more product entities. It goes beyond finding the polarity of the feature of an entity, but involves measuring the relation between the entities based on the sentiment keyword used. Considering a non-comparative sentence, “The camera is very expensive”, the polarity at sentence level can be determined by simply summing positive and negative sentiment keyword. In this case, the polarity is negative since there is only one keyword, “expensive” which has a negative sentiment score. For aspect level, similar analysis of sentiment keywords applies, but the keywords need to be linked to aspect/feature rather than sentence in general. For example, the word “expensive” is linked to “price” feature before the sentiment analysis takes place. In this case, sentiment for “price” feature will be negative.