In the modern era, there are massive amount of web resources present such as blogs, review sites and discussion forums. These resources form the platform where users can share their opinions or reviews` about anything whether it is a product, movie or a restaurant. Analysis of public sentiments deals with the determination of the polarity of different public opinions or reviews into either the category of positive, negative or neutral.
Thus, there comes the need of sentiment analysis which not only helps other individual to make a decision regarding buying a product, visiting a restaurant or watching a movie but also helps the producers of various products and owners of different restaurants to gain the knowledge of preferences of customers, so that it could be possible to increase the profit and economic value.
The paper presents a survey with main focus on performance of different artificial neural networks used for opinion mining or sentiment analysis while it also includes various machine learning approaches such as Naïve Bayes, Support Vector Machine, lexicon-based approach and Maximum Entropy.
ENTIMENT analysis is a kind of Natural Language Processing which includes the examining and analyzing of opinions regarding anything whether it is a product or a movie .
Thus, it is also referred as opinion mining. The opinions and reviews are posted online by the customers on different web resources that are available such as blogs or social media platforms . From such reviews, a person can easily determine whether the overall review and opinion of people concerning a particular thing is positive, neutral or negative.
But it would be a strenuous as well as a time consuming task for an individual to go through all the reviews regarding a particular restaurant or a movie prior to settling on a choice and making a decision. So, there comes the need and importance of classification of sentiments which could be performed at three different levels namely, document level sentiment analysis, sentence level sentiment analysis and sub-sentence level sentiment analysis which refers to examining the sentiments of a complete document, analyzing the reviews of a single sentence and classifying the opinion of the sub-expressions present in the sentence respectively .
There are two main approaches for analyzing the sentiments namely, Lexicon-based approach and the approach of machine learning. The techniques that come under the latter approach are Support Vector Machine (SVM), Naïve Bayes, Maximum Entropy and Artificial Neural Networks (ANN).
These techniques can be divided into two broad categories where Naïve Bayes and Maximum Entropy techniques come under probabilistic classifiers while Support Vector Machine and Artificial Neural Network form a part of linear classifiers.
This paper presents a survey on these different techniques with the main focus on the utilization of Artificial Neural Networks for the task of sentiment classification.
The remainder of this survey paper is organized as follows: Section II provides a literature review on sentiment analysis and Section III elucidates common methodology adopted in research papers. Further, Section IV presents the result of analysis of several research papers surveyed in both the sub-sections. Finally, Section V concludes the study and discusses the future scope for further research in the topic of sentiment analysis.
The first part of this survey paper, which is sub-section 1, comprises the papers that include the comparison of the techniques that come under supervised learning such as Support Vector Machine, Naïve Bayes and Maximum entropy.
Furthermore, the second sub-section includes the survey of the research papers that incorporate the artificial neural network models for opinion mining and sentiment classification and the comparative analysis of artificial neural network with other techniques.
In the year 2006, Hang et al. presented an idea of performing the process of sentiment analysis through the experimentation with different sets of algorithms that come under the category of machine learning, where feature selection was also conducted in order to achieve dimensionality reduction. Three algorithms were employed on online product reviews namely, discriminative classifier (Passive Aggressive (PA) algorithm), Winnow and generative model.
For discriminative classifier, PA algorithm was chosen because it followed an online learning pattern and it had a predictable performance. In a like manner, Winnow also adopted an online learning algorithm which predicted the polarity of the reviews. While the generative model was based on language model which was used to state the probability with which a given sequence of words can be generated and it used the ratio between positive and negative probabilities to give the polarity. It was depicted that the performance of the classifier was improved by high order n-grams.
Moreover, for the mixed reviews, discriminative model was proved to be efficient as it gave better result than other two models. The only drawback was that classification was only limited to just positive and negative scales and did not include the classification of reviews at different scales. Later, Qiang et al. proffered an experiment to classify the reviews that were associated with the blogs related to travelling for the seven destinations in 2009.
The sentiments were classified either into positive reviews or negative reviews, so that, it could be helpful to the both travelers, and the managers before taking any decision related to choosing the destination. Three approaches of machine learning were compared namely, Support Vector Machine (SVM), Naïve Bayes and a character based N-gram model which was a dynamic language model classifier. For the analysis by Naïve Bayes, a tool was developed by VC 6.0 and for the purpose of feature selection then the method of Information Gain was applied.
The SVM approach used the concept of word frequency whereas the N-gram based character language model preferred characters to be its basic unit. The dataset for the classification was the retrieved corpus from Yahoo.com and a Lingpipe DynamicLMClassifier was used for the experiment to be performed by the three algorithms. The k-fold cross validation was performed where three was chosen as the value of k. For the experiments, all the characters were converted into lowercase while the punctuation in itself was considered as a lexical item.
From the experiment, it was concluded that the performance of Support Vector Machine and N-gram was better in comparison to the performance of Naïve Bayes. Also, the performances of all the three approaches were significantly equal with large training dataset and had reached the 80% level of accuracy. The limitation of this approach was that only reviews regarding western countries were taken into consideration. Further, in 2011, Shi et al. came out with a technique of classifying the sentiments of hotel reviews on the basis of supervised learning.
Two types of features were considered for the approach of supervised learning using unigram feature which were, frequency and TF-IDF which means Term Frequency- Inverse Document Frequency. The feature described latter was a weight which basically measured the significance of a word in the whole document. The dataset, which was included, contained 4000 reviews which were divided into four folds of equal size that maintained a balanced distribution of the class.
The document was segmented using ICTCLAS and precision, recall and F-score were the three parameters for analyzing the performance. The experiment resulted in the realization that TF-IFD outperformed frequency in sentiment classification. Subsequently in 2011, the two approaches for sentiment classification namely, machine learning approach and the approach of lexicon-based look-up table were incorporated into a combined approach by Fang et al.
The dataset of camera reviews was chosen for sentiment classification and the classification was done on domain specific reviews and different expressions where sentiments were related to the specific domain. In order to implement the proposed system, three steps were followed where the first step was corpus filtering in which the phrases were extracted and an initial lexicon list was prepared. Secondly, the step of web search with linguistic pattern was performed with two patterns (camera aspect and seed words) which was followed by the expansion of dictionary with synonyms and antonyms.
Thus, two classifiers were combined which were, Aspect classifier and Polarity classifier. Four experiments were conducted out of which the experiment performed with incorporation of domain specific lexicon and knowledge encoded in MPQA in SVM approach gave the best accuracy of 47.4%. Likewise, Dhande and Patnaik in 2014 also introduced a classifier which incorporated a Naïve Bayes classifier so that it could analyze the reviews of several movies and classify them in accordance to their polarity as positive or negative.
The inputs that were selected were the movie review dataset and a keyword dictionary called the WordStat dictionary. Then, the reprocessing of the data was done using the model called Bag of Words (BOW) which used occurrence of each word and unigram feature in order to conduct the process of classification. There were two inputs of sentiment analysis namely the testing file and trained Naïve Bayes classifier following which the composition matrix was evaluated that contained the data about predicted and actual classification.
The experiment included simulation environment (MATLAB R2012a environment) and performance metrics (accuracy parameter). From the outcomes, it was inferred that the focused model came out to be the best with accuracy of 80.65%. After that, in 2015, Moh et al. proposed a multi-tier architecture for the purpose of classification where the sentiment analysis was performed on a movie review dataset which was extracted from rotten tomatoes database.
The introduced architecture focused on four different classifiers of supervised machine learning namely, SVM, Naïve Bayes, Random Forest and SGD (Stochastic Gradient Descent). The classification was done so as to categorize the sentiments into five different classes from very negative to very positive. The data was collected and cleaned for the purpose of pre-processing where the splitting of the train set was done according to 80-20 rule which was followed by feature selection.
Then, prediction was done using three models. Model-1 used the whole review dataset with three labels while Model-2 and Model-3 considered negative-very negative and positive-very positive labels respectively. The experiments were performed which resulted that multi-tier model outperformed single-tier model. Moreover, the SGD classifiers with Scikit learn performed significantly well using the dictionaries with an accuracy of 87.23%.
The approach of neural network was implemented to classify the reviews or sentiments of the people from blogs: Live Journal and Review Centre along with movie dataset by Chen et al. in 2011 where the concept of semantic orientation was applied and incorporated with machine learning for effective sentiment classification. The back propagation neural network (BPN) was used along with four different indexes for sentiment orientation. The first index was SO-A index that infer the semantic orientation from the association where a paradigm was prepared and strength was calculated.
Second and third indexes were SO-PMI (AND) and SO-PMI (NEAR) which used Point wise Mutual Information and Information Retrieval (PMI-IR) while the last index was SO-LSA. The proposed methodology was to prepare a dataset to form Term Document Matrix (TDM) and calculate its’ SO-indexes. Then the network was trained and the performance was evaluated which concluded that the method performed better than several traditional techniques and was also succeeded in reducing the time for the classification significantly. The potential limitation of the proposed method was that it incorporated only blogs while twitter, face-book could also have been used.