Over the last decade industrial and academic communities have increased their focus on sentiment analysis techniques, especially applied to tweets. State-of-the-art results have been recently achieved using language models trained from scratch on corpora made up exclusively of tweets, in order to better handle the Twitter jargon. This work aims to introduce a different approach for Twitter sentiment analysis based on two steps. Firstly, the tweet jargon, including emojis and emoticons, is transformed into plain text, exploiting procedures that are language-independent or easily applicable to different languages. Secondly, the resulting tweets are classified using the language model BERT, but pre-trained on plain text, instead of tweets, for two reasons: (1) pre-trained models on plain text are easily available in many languages, avoiding resource- and time-consuming model training directly on tweets from scratch; (2) available plain text corpora are larger than tweet-only ones, therefore allowing better performance. A case study describing the application of the approach to Italian is presented, with a comparison with other Italian existing solutions. The results obtained show the effectiveness of the approach and indicate that, thanks to its general basis from a methodological perspective, it can also be promising for other languages.

https://www.mdpi.com/1424-8220/21/1/133/pdf

An Effective BERT-Based Pipeline for Twitter Sentiment Analysis: A Case Study in Italian.

Social media offer a big amount of information, to exploit in many fields of research. However, while methods for Natural Language Processing are being developed with good results when applied to well-formed datasets made of written text with a clear syntax, these sources present text written in informal language, unstructured syntax, and with peculiar symbols; therefore, particular approaches are required for text processing in this case. In this paper, the task of sentiment analysis of tweets is regarded. In particular, in order to avoid noise constituted by some web constructs like URLs and mentions and by other text fragments, and to exploit information hidden in symbols like emoticons, emojis and hashtags, the pre-processing of tweets is analyzed. More in detail, a number of experiments, performed by a state-of-the-art classification model (BERT), are designed, to evaluate many currently available operations for pre-processing tweets, in terms of the statistical significance of their influence on sentiment analysis performances. Moreover, available data in two languages are considered, i.e., English and Italian, in order to also evaluate dependence on the language. Results allow to individuate the most convenient strategy to pre-process tweets, and thus to improve the state of the art in both languages for the considered task of sentiment analysis.

Multilingual evaluation of pre-processing for BERT-based sentiment analysis of tweets

The analysis of sentiments is essential in identifying and classifying opinions regarding a source material that is, a product or service. The analysis of these sentiments finds a variety of applications like product reviews, opinion polls, movie reviews on YouTube, news video analysis, and health care applications including stress and depression analysis. The traditional approach of sentiment analysis which is based on text involves the collection of large textual data and different algorithms to extract the sentiment information from it. But multimodal sentimental analysis provides methods to carry out opinion analysis based on the combination of video, audio, and text which goes a way beyond the conventional text‐based sentimental analysis in understanding human behaviors. The remarkable increase in the use of social media provides a large collection of multimodal data that reflects the user's sentiment on certain aspects. This multimodal sentimental analysis approach helps in classifying the polarity (positive, negative, and neutral) of the individual sentiments. Our work aims to present a survey of recent developments in analyzing the multimodal sentiments (involving text, audio, and video/image) which involve human–machine interaction and challenges involved in analyzing them. A detailed survey on sentimental dataset, feature extraction algorithms, data fusion methods, and efficiency of different classification techniques are presented in this work.

Multimodal sentimental analysis for social media applications: A comprehensive review

Data analytics and its associated applications have recently become impor-tant fields of study. The subject of concern for researchers now-a-days is a massive amount of data produced every minute and second as people con-stantly sharing thoughts, opinions about things that are associated with them. Social media info, however, is still unstructured, disseminated and hard to handle and need to be developed a strong foundation so that they can be utilized as valuable information on a particular topic. Processing such unstructured data in this area in terms of noise, co-relevance, emoticons, folksonomies and slangs is really quite challenging and therefore requires proper data pre-processing before getting the right sentiments. The dataset is extracted from Kaggle and Twitter, pre-processing performed using NLTK and Scikit-learn and features selection and extraction is done for Bag of Words (BOW), Term Frequency (TF) and Inverse Document Frequency (IDF) scheme. 
For polarity identification, we evaluated five different Machine Learning (ML) algorithms viz Multinomial Naive Bayes (MNB), Logistic Regression (LR), Decision Trees (DT), XGBoost (XGB) and Support Vector Machines (SVM). We have performed a comparative analysis of the success for these algorithms in order to decide which algorithm works best for the given data-set in terms of recall, accuracy, F1-score and precision. We assess the effects of various pre-processing techniques on two datasets; one with domain and other not. It is demonstrated that SVM classifier outperformed the other classifiers with superior evaluations of 73.12% and 94.91% for accuracy and precision respectively. It is also highlighted in this research that the selection and representation of features along with various pre-processing techniques have a positive impact on the performance of the classification.  The ultimate outcome indicates an improvement in sentiment classification and we noted that pre-processing approaches obviously suggest an improvement in the efficiency of the classifiers.

/pdf/influence-of-pre-processing-strategies-on-the-performance-of-18xaz9at2d.pdf

Influence of Pre-Processing Strategies on the Performance of ML Classifiers Exploiting TF-IDF and BOW Features

The emissions of greenhouse gases, such as carbon dioxide, into the biosphere have the consequence of warming up the planet, hence the existence of climate change. Sentiment analysis has been a popular subject and there has been a plethora of research conducted in this area in recent decades, typically on social media platforms such as Twitter, due to the proliferation of data generated today during discussions on climate change. However, there is not much research on the performances of different sentiment analysis approaches using lexicon, machine learning and hybrid methods, particularly within this domain-specific sentiment. This study aims to find the most effective sentiment analysis approach for climate change tweets and related domains by performing a comparative evaluation of various sentiment analysis approaches. In this context, seven lexicon-based approaches were used, namely SentiWordNet, TextBlob, VADER, SentiStrength, Hu and Liu, MPQA, and WKWSCI. Meanwhile, three machine learning classifiers were used, namely Support Vector Machine, Naïve Bayes, and Logistic Regression, by using two feature extraction techniques, which were Bag-of-Words and TF–IDF. Next, the hybridization between lexicon-based and machine learning-based approaches was performed. The results indicate that the hybrid method outperformed the other two approaches, with hybrid TextBlob and Logistic Regression achieving an F1-score of 75.3%; thus, this has been chosen as the most effective approach. This study also found that lemmatization improved the accuracy of machine learning and hybrid approaches by 1.6%. Meanwhile, the TF–IDF feature extraction technique was slightly better than BoW by increasing the accuracy of the Logistic Regression classifier by 0.6%. However, TF–IDF and BoW had an identical effect on SVM and NB. Future works will include investigating the suitability of deep learning approaches toward this domain-specific sentiment on social media platforms.

/pdf/climate-change-sentiment-analysis-using-lexicon-machine-1htytylx.pdf

Climate Change Sentiment Analysis Using Lexicon, Machine Learning and Hybrid Approaches

In the big data era, data is made in real-time or closer to real-time. Thus, businesses can utilize this evergrowing volume of data for the data-driven or information-driven decision-making process to improve their businesses. Social media, like Twitter, generates an enormous amount of such data. However, social media data are often unstructured and difficult to manage. Hence, this study proposes an effective text data preprocessing technique and develop an algorithm to train the Support Vector Machine (SVM), Deep Learning (DL) and Naive Bayes (NB) classifiers to process Twitter data. We develop an algorithm that weights the sentiment score in terms of weight of hashtag and cleaned text. In this study, we (i) compare different preprocessing techniques on the data collected from Twitter using various techniques such as (stemming, lemmatization and spelling correction) to obtain the efficient method (ii) develop an algorithm to weight the scores of the hashtag and cleaned text to obtain the sentiment. We retrieved N=1,314,000 Twitter data, and we compared the popularity of two products, Google Now and Amazon Alexa. Using our data preprocessing algorithm and sentiment weight score algorithm, we train SVM, DL, NB models. The results show that stemming technique performed best in terms of computational speed. Additionally, the accuracy of the algorithm was tested against manually sorted sentiments and sentiments produced before text data preprocessing. The result demonstrated that the impact produced by the algorithm was close to the manually annotated sentiments. In terms of model performance, the SVM performed better with the accuracy of 90.3%, perhaps, due to the unstructured nature of Twitter data. Previous studies used conventional techniques; hence, no precise methods were utilized on cleaning the text. Therefore, our approach confirms that proper text data preprocessing technique plays a significant role in the prediction accuracy and computational time of the classifier when using the unstructured Twitter data.

Saurav Pradha

Papers

Effective Text Data Preprocessing Technique for Sentiment Analysis in Social Media Data