States and Canada [28]. Sentiment grabber model was developed using Ontology, probabilistic LDA and text annotations [13]. Hotel reviews were automatically classified by using SVM and fuzzy domain ontology [29].
Research on user-generated content was also focused on the lexicon-based or linguistic-based approach. Named entity recognition, feature extraction, reliability of content, language used are some of the challenges exist for text analysis. Information Retrieval (IR) techniques like Vector Space Modelling, Latent Semantic Indexing (LSI) and Latent Dirichlet Allocation (LDA) were used for transforming unstructured free text into structured format. Words which do not represent entities were removed from consumer product reviews by using PMI measure, to improve the precision of feature extraction method [30]. In the lexicon-based approach, positive and negative words were extracted from the opinions [31–34], and overall sentiment aggregation was determined for the documents. Words or phrases in the presence of conjuncts and connectives were considered to build word dependencies. Sentiment analysis was then done using Naive Bayes classification algorithm [31, 34]. Opinion observer was built using NLP techniques for detecting the polarity of opinions and by using the opinion aggregation function [35]. Automatic extraction of adjectives related to sentiments from blogs and reviews was proposed and used association rule mining for building the dictionary [32], which resulted in the accuracy of more than 70% for positive adjectives and more than 60% for negative adjectives. Similarly, sentiment dictionaries were created using naive Bayes algorithm and NLP techniques for developing opinion mining model for film reviews [36]. Poirier et al. concluded that machine learning algorithms were suitable for larger data set, whereas linguistic methods were suitable for smaller data set.
Double propagation method was proposed for the retrieval of new sentiments from sentences and positive or negative polarity was assigned for them [33]. Product’s features were extracted using unsupervised learning techniques [11] from the review documents, and words belong to the same concept are grouped using Latent Semantic Association (LaSA) model. Text analysis and statistical techniques were used to rank the product quality from their websites [31]. NLP techniques were used to identify the most frequently used positive and negative sentiment words for the classification of movie documents [37]. Non-negative matrix factorization and clustering techniques were used for retrieving suitable answers for the given query as a text summarization technique [38]. Lexicon-based NLP techniques were used to extract conjunctions, connectives, modals and conditionals for sentiment polarity detection of tweets [34]. Basiri et al. [39] used Dempster–Shafer theory for sentiment aggregation at document level using the mass function. The probabilistic based Latent Dirichlet Allocation (LDA) was used for annotation of semantics in text documents [13].
The user-generated content, which are in unstructured or semi-structured format, can be converted into structured format using NLP and machine learning techniques, and is made available for decision-making purposes. Multi-Criteria Decision Making (MCDM) techniques are used in different sectors like in fast food restaurants for measuring service quality [40], for ranking universities [41] and in different simple and complex industrial applications [42–45]. Customer lifetime value and their loyalty were evaluated based on the hybrid approach by combining Analytic Hierarchy Process (AHP) and association rule mining [46]. The best alternative for oil project fields was evaluated using AHP for weights identification and fuzzy TOPSIS for ranking process [44] and as the service quality indicators for tourism industry in Iran [47]. Different MCDM techniques along with statistical techniques were applied in different sectors like healthcare sector [48], movie recommender systems [49] for its performance measurement so as to improve its quality of services. MCDM technique like VIKOR was used for the measurement of customer satisfaction and ranking of mobile services [50] and for ranking the suppliers [51].
Ontology learning includes extraction of domain terms from the sources, modelling of data through Ontology development and easy retrieval while querying. Manual building of Ontology takes greater effort and it is complex and challenging. This motivates the researchers to automatically generate Ontology for the domain specific terms present in the social media reviews written for a product/service. The Ontology-based Semantic Indexing (OnSI) method tags concepts and attributes, into the Ontology using the contextually related words. It enables query processing and further information retrieval processing easier in subsequent steps. This semantic-based approach of indexing improves higher accuracy while identifying the concepts or attributes (or features) from the contents of text documents [26, 27]. Ontology-based approach for mobile product review classification was resulted in precision 75% and in recall 40% [52], and recall more than 82% [27].
1.3 Motivation
Feature extraction from product or service review documents often includes different steps like data pre-processing, document indexing, dimension reduction, model training, testing, and evaluation. Labeled data set of document collection is used to train or learn the model. Further, the learned model is used for identifying unlabeled concept instances from the new set of documents. Document indexing is the most critical and complex task in text analysis. It decides the set of key features to represent the document. It also enhances the relevancy between the word (or feature) and the document. It needs to be very effective as it decides the storage space required and query processing time of documents.
The Ontology-based or semantic-based approach is used to retrieve the concepts from the documents by establishing the contextual relationships. In content-based approach, BagOfWords model is used for representing the text, where synonymy and polysemy cannot be resolved as they use terns as indexes. However, in context-based semantic approaches like topic modeling techniques, concepts are used to extract information and their categorization. It projects the contextual relationship among the terms present in the documents.
In order to utilize the strength of Ontology in user-generated content analysis process, this chapter proposes a domain Ontology-based Semantic Indexing (OnSI) technique for product or service reviews generated by the customers in the social media platform. The integration of topic modelling technique with the Ontology learning is explained by the OnSI method; it is generic and is applicable to any domain.
1.4 Feature Extraction
The World Wide Web has large amount of text documents and the necessity of annotating them has become vital. Customers would express their views by writing product reviews, user feedback in the descriptive format, which is mostly in unstructured. This format makes difficult for the machines to process these text documents. Hence, it becomes necessary to annotate large volume of text in order to develop business intelligence or automated solutions. These data have to be analyzed and modelled for enabling the decision-making process. The challenging task of extracting the information is made easier by adding, annotating documents, which in turn paves the way for automated solutions [53].
Feature extraction is the process of building dataset with informative and non-redundant features from the initial set of data. The subsequent methods like feature selection reduces the amount of resources required for its representation. Many machine learning algorithms like classification and clustering are used for extracting the features such as entities and attributes from the text documents using their properties which are similar. The challenges like absence of semantic relations between entities while feature selection and lack of prior knowledge in domain may be overcome by applying suitable NLP and IE techniques [54, 55].
Feature extraction from product or service review documents often includes different steps like data pre-processing, document indexing, dimension reduction, model training, testing, and evaluation. Labeled data set of document collection is used to train or learn the model. Further, the learned model is used for identifying unlabeled concept instances from the new set of documents. Document indexing is the most critical and complex task in text analysis. It decides the set of key features to represent the document. It also enhances the relevancy between the word (or feature) and the document. It needs to be very effective as it decides the storage space required and query processing time of documents.
Pre-processed data is built into Term Document (TD) matrix by using term weighing schemes, as shown in Figure