This section provides an overview of the most relevant topics related to text sentiment analysis. First the definition of Online Consumer Reviews (OCRs) is provided; then, sentiment analysis-related topics were identified and discussed.
Online consumer reviews (OCRs)
OCR is one of the most commonly used concepts to represent the traditional word-of-mouth review. An electronic word-of-mouth review, or OCR, is defined as “any positive or negative statement made by potential, actual, or former customers about a product or company, which is made available to a multitude of people and institutions via the Internet” [24]. Word-of-mouth review, meaning personal opinions among people, has been recognized as a significant source of information to understand customers’ interests, and sentiments concerning companies’ products and services, such as movies, books, music albums, and enterprises such as hotels, and restaurants. Many consumers find word-of-mouth information to be useful and credible when making a decision about products or services because it is generated by independent pre-experienced consumers instead of biased company advertisements. With the rapid advancement of Internet technology, the electronic word-of-mouth technique has been adopted by different platforms such as Yelp, Amazon, and eBay to enable people to easily generate reviews, share them with other people, and exchange opinions. Electronic word-of-mouth information includes customer reviews, online comments, and score ratings, and it can be spread in real-time through online channels, such as e-commerce sites, online forums, the blogosphere, and social networking sites. Thus, electronic word-of-mouth information is recognized as not only a convenient way for consumers to share information, but also a source of new challenges and opportunities for business analysts to understand consumer interests and opinions.
Statistics also support reliance on Online Consumer Review (OCR) for the decision-making of consumers. Nearly 65% of consumers access consumer-written product reviews via the Internet [15]. Additionally, of those consumers who read reviews, 82% confirmed that reviews had directly influenced their buying decision, while 69% shared the reviews with others including: family, friends, and co-workers, so magnifying their impact. In addition, numerous surveys and consulting reports have suggested that, for a number of consumers and products (not applicable to all), consumer-generated reviews are valued more highly than reviews from ‘experts’ [21, 61]. Therefore, Online Consumer Reviews (OCRs) can impact the consumer decision-making process to a greater extent than traditional media [1].
Text mining and sentiment analysis
Data and text mining cover the broad scope of software tools and mathematical modelling approaches which are used to discover implicit, previously unknown patterns from data. In text mining, patterns are lifted from natural language text (unstructured data), while, in data mining, the patterns are lifted from structured databases. The text mining process starts with the text collection stage and then proceeds to the pre-processing stage in which the text is cleaned and formatted. The pre-processing stage involves critical tasks, such as tokenization, removal of stop words, and stemming. In the next stage, meaningful features are extracted to make inferences about the data. In the final stage, text mining approaches, such as categorization, topic modelling, or clustering, are applied to answer certain questions about the given data.
Large volumes of Online Consumer Reviews (OCRs) are available on many retailers' websites, and mining such data to understand consumers’ opinions is called opinion mining. This term was first coined by Dave et al. [14], where opinion mining involves processing a set of reviews for a given product or service and extracting a list of attributes or aspects to categorize the consumer opinion into different classes such as (positive, negative, subjective, objective, etc.). Sentiment analysis is a subsection of opinion mining which focuses on the extraction of the consumer's emotions, opinions, and evaluations on services or products from online reviews they have posted. Sentiment analysis is an area of research that is very active, with a large volume of relevant research literature available [3]. Most of the research work is typically based on three common sentiment analysis tasks: subjectivity analysis, polarity detection, and sentiment strength detection. Subjectivity analysis aims to determine whether or not a given text is subjective, while polarity detection is utilized to assign an overall positive or negative sentiment orientation to subjective texts. Sentiment strength detection specifies the degree of polarity to which a text is either positive or negative.
Sentiment analysis is normally performed in two ways: a lexicon-based approach or a machine learning approach. A lexicon-based method uses a sentiment dictionary or a sentiment lexicon that is used to predict the overall sentiment of a text based on pre-defined word occurrence. Alternatively, a machine learning approach generates a classifying algorithm through learning with the set of linguistic features [28]. The trained classifier is then used for sentiment prediction [3].
In the lexicon-based approach, public lexicons, such as SenticNet [10], SentiWordNet [2], and OpinionFinder [63], have been frequently applied by many studies owing to the reliability of public sentiment dictionaries [28]. Lists of sentiment-bearing words and phrases available in opinion lexicon are used for lexicon-based techniques, such as the General Inquirer lexicon [54], WordNet Affect [55] SentiWordNet, the ANEW words [8], and the LIWC dictionary [43]. Beyond these standard resources and to automatically generate and score lexicons, researchers have created new methods. However, as indicated by Liu Y [34], while an opinion lexicon is required, it is insufficient for sentiment analysis. Thus, a combined approach is more appropriate as these approaches normally use additional information, e.g. semantic rules to handle emoticon lists, negation, booster word lists, and an already existing and substantial collection of subjective logical statement patterns. According to Taboada et al. [56], “lexicon-based methods for sentiment analysis are robust, result in good cross-domain performance, and can be easily enhanced with multiple sources of knowledge.”
LEX-Quality of Experience (QoE) parameters were utilized by [61] to analyse user reviews. The identification of frequent nouns in reviews was achieved through the utilization of speech tagging, and these were denoted as a prospective QoE element. Semantic lexicons, such as SentiWordNet, were used to group and aggregate similar nouns. For each group, the representative nouns were highlighted as QoE parameters. This work, therefore, exploited user reviews as inputs for quality element extraction from services through the selection of frequent nouns in drawing features and the end outcome.
Recently, machine learning algorithms have been used for most existing sentiment analysis techniques, such as Naive bayes (NB), support vector machines (SVM), neural network (NN), genetic algorithm (GA), and k-nearest neighbours (kNNs) to optimize, classify, and form predictions based on the data in text documents. Machine learning approaches have certain advantages, including the ability to identify the non-sentiment terms which express a sentimental judgement (e.g. “cheap” in the phrase “this camera is cheap”). An additional advantage of such approaches is the availability of a wide range of applicable learning algorithms. However, these methods present certain disadvantages, such as the need for a human-labelled corpus for the training phase. Additionally, while within the domain these trained machine learning methods perform very well, their performance can diminish significantly when applied to another domain. For example, in the cell phone domain, the words “cheap” and “smart” are used as expressions of positive opinions, while in the world of books domain, “well-researched” and “thriller” signify positive sentiments. Therefore, a cell phone domain-trained algorithm is unlikely to correctly classify book domain reviews. Moreover, as indicated by [58], some machine learning algorithms cannot “give a clear explanation as to why a sentence has been classified in a certain way, by reference to the predefined list of sentiment terms.”
One can principally investigate sentiment analysis applications at three granularity levels: document level, sentence level, and aspect level. At the document level, the entire document is allocated an overall sentiment score. Sentence-level sentiment analysis concentrates on predicting the sentiment of stand-alone sentences. Subsequently, a score aggregation method is applied to generate an overall review score from combined sentence-level scores. However, in a document- or sentence-level analysis, it is not easy to obtain fine-grained opinions, though an aspect-level analysis can frequently overcome this problem. Aspect-level techniques carry out a finer-grained analysis with the intention of identifying sentiments on entities and/or their aspects [65].
Challenges in using sentiment analysis with OCRs
There are certain challenges and problems in implementing sentiment analysis, and some of them are as follows:
Short reviews: in a proposal by Cosma et al. [13] they state that in order to surmount the domain barrier in gathering views, there is need for an overall way of setting up language rules for the recognition of view-bearing words. Additionally, online reviews have unique text features that are short in length, use formless phrases, and involve substantial data. New challenges to conventional study topics in text analytics, i.e. text categorization, data mining, and emotional studies, are brought to the short reviews.
Colloquial language is another vital attribute of online text, specifically in online reviews. Consumers may use short forms or acronyms that rarely appear in traditional text when writing reviews, for example, phrases like “superb”, “good 2go”, hence making it extremely hard for one to identify the semantic meaning [5].
Mockery acknowledgement: the varied challenges require working through mockery or expressions that are unexpected. Riloff et al. [44] contributed to the improved review in mockery acknowledgement by developing an algorithm that naturally learned to group good and unpleasant phrases for tweets. The evaluation of two elements that are dissimilar amounts to analogy.
Domain dependency: the essential task of exploring the information generated by the customer lies in the wider concept of themes. Generally, the content generated is usually broad and needs to be packaged into categories. A classifier that is specified for a given domain space might thus not be effective in another domain which uses different words. The expression of sentiments is varied in different domains. This, notwithstanding the methods of sentiment categorization, can be synchronized to adequately work in a provided domain but, at the same time, may be limited to categorize sentiments in a varying domain. In light of this Bollegala et al. 2013 [7] proposed a cross-domain sentiment classifier that automatically extracts a sentiment thesaurus. Moreover, procedures or algorithms that are joined in a given area may not necessarily perform effectively in a space that is different from the initial one. The process of identifying specific area- and space-free systems was independently carried out. Jambhulkar and Nirkhi [25] carried out a cross-domain sentiment analysis survey study that focused on the following methods: sentiment-sensitive thesaurus, spectral feature alignment, and structural correspondence learning. The findings of the study denoted that each of the used methods has its distinct way in (1) increasing the vector features, (2) evaluating the association between given words, as well as (3) the used classifier. According to Bisio et al. [4] there are two main features of notion characterization. These include the versatile nature of a provided structure and the subsequent ability to work in wider business spaces through utilization of relevant valence shifters, semantic systems, and a predictive model grounded on distance.