Sentiment Analysis & Machine Learning Techniques

Artificial intelligence (AI) / Machine learning (ML) techniques are getting more and more popular. Many people use machine learning to analyze the sentiment of tweets, for example, to make predictions related to different business areas. In this blog post, you will learn about different machine learning / deep learning and NLP techniques which can be used for sentiment analysis.

What is sentiment analysis?

Sentiment analysis is about predicting the sentiment of a piece of text and then using this information to understand users’ (such as customers) opinions. . The principal objective of sentiment analysis is to classify the polarity of textual data, whether it is positive, negative, or neutral. Whether the end-user sentiment is positive or negative or neutral can be used to answer many different business questions. The text whose sentiment needs to be processed can be extracted from many different sources, but in the current scenario, sentiment analysis is mostly about tweets and reviews.

The unprecedented abundance of data available on the internet on different social media websites has attracted business and research interest from various different fields including marketing, political science, and social studies, etc. The following are some example questions that can be dealt with using sentiment analysis.

• Do people like the new OnePlus phone?
• What do people hate about Oneplus latest phone?
• What do people think about the current visit of India’s prime minister to the US?
• What do people think of the US exit from Afghanistan?
• How do we recognize the emergence of health problems such as depression and cancer?
• What do people think about the latest budget presented by the Indian government?

People employ three different modalities to communicate in a coordinated manner. They are the following:

• The language modality with the use of words and sentences
• The visual modality with gestures, poses, and facial expressions. This is termed visual sentiment analysis.
• The acoustic modality through the change in vocal tones. This can be termed voice sentiment analysis.

Sentiment analysis would require the representation of all of the above modalities to estimate the human sentiment in the most accurate manner. Multimodal representation learning has shown great progress in a large variety of tasks including emotion recognition, sentiment analysis. Multimodal sentiment analysis is a trending area of research, and multimodal fusion is one of its most active topics.

One form of sentiment analysis is aspect-based sentiment analysis (ASBA).  Aspect-based sentiment analysis is a task in which the sentiment for each aspect of an entity is determined. Aspects can be a feature, a characteristic, or behavior of a product or an entity, such as the ambiance of a restaurant, the performance of a laptop, the display of a phone, and so on. Customer feedback about the aspects can help manufacturers and merchants develop ways including modifying products and services offerings in order to increase customer happiness. Thousands of customer reviews covering various aspects and their corresponding opinions can be found on the review pages/sections of any product.

Another form of sentiment analysis is visual sentiment analysis. There is increasing attention in visual sentiment analysis driven by the need for more and more people to share their feelings with images, emojis, and other visual content. Visual sentiment analysis is defined as a machine learning task to classify whether a given image shows a positive, negative or neutral sentiment. The visual sentiment is related to social media marketing and customer feedback analysis. It’s been utilized in many areas, including advertising and market research, where photos are shared on a variety of internet sites, such as Facebook, Twitter, Instagram, and others., for the purpose of promotion of products/services and gathering feedback from users.

Different machine learning techniques to use for sentiment analysis

Sentiment analysis can be done using techniques related to natural language processing (NLP) and machine learning. NLP techniques such as bag-of-words (BoW) and term frequency-inverse document frequency (TF-IDF) can be used. Machine learning algorithms such as Support Vector Machine (SVM), Logistic Regression, Multinomial Naive Bayes, Random Forest, artificial neural networks (ANN), deep learning techniques such as LSTM, bi-directional LSTM etc.

NLP techniques are used to pre-process and vectorize the text data. The vectorized data is then used for training different machine learning models based on different algorithms. The following is the list of different steps required to train the model for sentiment analysis:

• Collect dataset to train and test machine learning (ML) model classifier.
• Pre-process the dataset for subsequent processing. The following are some of the steps which are done as part of text preprocessing:
  • Stopword removal
  • Non-standard (slang) to standard word mapping
  • PoS tagging
  • Tagging positive/negative words
  • Stemming
• Convert textual data into vector form using NLP techniques such as BoW or TF-IDF.
• Divide the dataset into training and testing groups.
• Train the ML classifier with training data. Use algorithms such as SVM, logistic regression, multinomial Naive Bayes, random forest, etc.
• Predict the polarity of testing data
• Use evaluation metrics such as accuracy, precision, recall, and F1-score for model evaluation
• Perform algorithm selection and model selection steps to select the best model.

For training the machine learning models, the text data would need to be manually labeled. For example, in sentiment analysis related to tweets and reviews, machine learning models are trained with labeled data of sentiments.

In visual sentiment analysis, machine learning is used to classify the sentiments of images based on their colors, texture, shape, etc. A machine learning approach first learns about different types of inputs (images with different sentiments), and then uses this information to label new unlabeled data as either positive or negative sentiment. With the successes of deep neural networks (CNN) in conventional computer vision tasks, numerous methods have been proposed to conduct visual sentiment analysis and have shown clear advantages over traditional methods with handcrafted features. There are, however, challenges related to visual sentiment analysis such as data labeling. The data labels for visual sentiment analysis are inherently subjective and error-prone since it can be confusing for humans to recognize the sentiment of images.

Real-world examples of machine learning applications and sentiment analysis

The following represents some real-world examples of machine learning applications that use machine learning to analyze the sentiment of texts:

• Sentiment analysis in social media: Social networks such as Twitter, Facebook, etc are some of the most popular places for people to express their opinions about different topics. It is important for marketers to understand what customers or end-users think about a product or service because this information can be used to market their campaign and help their business grow. In the case of Twitter, the sentiment analysis is used to determine the average sentiment of a large number of tweets related to a specific product or service, and then this information can be analyzed by marketing experts in order to make decisions about what strategy they need to implement for their future marketing campaigns.
• The sentiment analysis of customer reviews: The product or service review section on different e-commerce websites is another machine learning application where machine learning can be used to analyze the sentiment of users’ opinions about a product. Companies use this information to improve future versions of their products and services, better understand customers, etc. Regarding customer reviews, there can be overall sentiments, and then, there can be aspect-based sentiments. And, one might need to train different machine learning models to predict overall sentiment and aspect-based sentiments. The aspect-based sentiment analysis problem has two sub-problems: 1) aspect extraction representing features of the product & service and 2) finding the sentiment/polarity toward each aspect. And, the aspect extraction involves two sub-tasks: 1) extracting aspect terms and 2) categorizing/normalizing the extracted aspect terms into aspect categories. Learn more about aspect-based sentiment analysis from this paper: Exploring conditional text generation for aspect-based sentiment analysis.
• Sentiment analysis for stock market prediction: The stock prices of different companies show a lot of fluctuations and machine learning models can be used to predict whether the stock price would go up or down in certain time intervals. Comments about specific stocks can be read from social media sites such as Reddit, Twitter, etc. Note that the review comments can be from investors and non-investors and this needs to be factored in while building the model. When it comes to sentiment analysis for predicting changes in stock prices, there are two machine learning tasks: A. Predicting the overall sentiment (positive/negative) about a stock (entity) in the news. This machine learning task is called entity sentiment analysis and this type of machine learning model can be useful for predicting stock prices based on certain events (hiring/firing employees, new products, etc). B. Predicting aspects related to a particular stock event or topic (for example: what would happen if Apple releases its new product). This machine learning task is called aspect-based sentiment analysis and the machine learning model for this type of prediction can be helpful in predicting stock prices.

References

• Sentiment Analysis in Twitter for Macedonian
• A Comparative Study of Sentiment Analysis Using NLP and Different Machine Learning Techniques on US Airline Twitter Data
• Learning to Aggregate and Refine Noisy Labels for Visual Sentiment Analysis
• Stock price prediction under anomalous circumstances
• Improving multimodal fusion via mutual dependency maximisation

In this blog post, we’ve discussed machine learning techniques that can be used for sentiment analysis. Whether you need to predict the polarity of a text or analyze customer reviews about your product, machine learning is an excellent way to do so. In order to train a machine learning model, it requires vectorizing textual data and dividing it into training and testing groups. Once that’s complete, there are many machine learning algorithms from which can be used to classify texts as positive or negative sentiment based on their features.