Sentiment Analysis

LIS 4/5693: Information Retrieval and Text Mining

Dr. Manika Lamba

What is Sentiment Analysis?

A computational process of identifying and categorizing opinions expressed in a piece of text, to determine whether the writer's attitude towards the topic is positive, negative, or neutral

• Analyze digital text to determine emotional tone of text

• Organizations have large volume of text:

  • Emails

  • Customer support chat transcripts

  • Social media comments

  • Reviews

• Tools can scan text automatically to determine writer’s attitude

Let’s start this lecture by defining sentiment analysis. Sentiment analysis is a computational process of identifying and categorizing opinions expressed in a piece of text to determine whether the writer’s attitude towards the topic is positive, negative, or neutral. These are the broad categories for sentiment analysis. You could, of course, have more granular categories if you wanted to, like very positive or mildly positive, and so on.

So what do you do when you perform a sentiment analysis?

You’ll analyze text available in digital form to determine the emotional tone of the message. Now, organizations have been collecting all kinds of text for all kinds of reasons for a long time. This text can include emails, customer support chat transcripts, social media comments, or even reviews. All of this text contains some information that will be useful to the organization. Sentiment analysis tools can automatically scan text to determine the writer’s attitude and tone, and this can give organizations important insights into their customers.

Importance of Sentiment Analysis

• Objective insights
• Enhanced product and service development
• Scalability in data analysis
• Real-time market response

So how can sentiment analysis tools help businesses? They offer businesses objective and consistent results in evaluating customer opinions, reducing personal bias that might occur with human reviewers. This allows organizations understand diverse customer perspectives and learn how customers respond to their products. These tools can also help organizations enhance their products and services. By identifying specific entities and sentiments associated with them, sentiment analysis enables companies to pinpoint areas for improvement in their products and services. These tools enable scalability in data analysis. There is a vast amount of structured data generated from various sources, emails, surveys, customer feedback.

Sentiment analysis tools allow businesses to efficiently process and analyze this data at scale, helping them stay informed about customer opinions and trends. And real-time sentiment analysis empowers businesses to quickly respond to emerging trends or potential crises. Businesses can receive instant alerts on negative sentiments and take proactive measures to address customer concerns.

Sentiment Analysis Use Cases

• Improve customer service
• Brand monitoring
• Market research
• Track campaign performance
• Refine public relation strategies
• Product or service monitoring
• Support political analysis

So where could businesses use sentiment analysis in improving their customer service?

This allows customer support teams to personalize their responses by understanding the mood of the conversation. By analyzing sentiments expressed in social media forums, blogs, and news articles, organizations can monitor their brand’s public perception.

Sentiment analysis can help understand customer preferences and dislikes based on their feedback on various platforms, and this assists in market research for new products and services. Marketers can use sentiment analysis to gauge the public reaction to their advertising campaigns. These tools can also be used for planning and implementing effective PR strategies, understanding customer grievances, identifying trends, and leveraging social media influencers.

Sentiment analysis can be used to gauge customer reactions to new products or services using feedback from e-commerce sites and other platforms. Sentiment analysis is very useful in the political arena as well. Public platforms like Twitter, now X, help authorities gauge public reaction to new policies, and political parties can use this data to refine their strategies.

Approaches to Sentiment Analysis

• Rule-based

• Machine Learning

• Training Machine Learning Model

• Using Model for Predictions

Let’s, discuss in a little more detail some of the common approaches to sentiment analysis which include: Rule-based sentiment analysis, Machine learning-based sentiment analysis, and hybrid sentiment analysis.

Let’s start by understanding how rule-based sentiment analysis works. This is where you’ll set up human-crafted rules to help identify subjectivity, polarity, or subject matter of the text. You’ll use some kind of system. Let’s say words like happy, affordable, fast will be in the positive lexicon, and words like poor, expensive, and difficult will be in the negative lexicon. You’ll then set up word scores for positive and negative words. Let’s say scores from 5 to 10 denote positive words, scores from -1 to -10 denote negative words. Thus, every document will generate a score, with higher scores for positive words and lower scores for negative words. So overall, if the document has a high score, you know it has positive sentiment.

A rule-based sentiment analysis system is straightforward to set up, but it’s hard to scale. It’s also a naive approach because it does not account for the positions of words in a sentence or a document. When you discover new keywords for conveying intent in the text input, you’ll need to keep expanding your lexicons. This approach may also not be accurate when processing sentences influenced by different cultures.

Next, let’s go to machine learning-based approaches to sentiment analysis. Sentiment analysis is, at its heart, a classification problem. You are trying to categorize text as positive, negative, or neutral sentiment, so you can use machine learning or deep learning techniques to build these classification models that are trained on large datasets, with diverse examples belonging to the different sentiment categories. This allows the software to differentiate and determine how different word arrangements affect the final sentiment score. Of course, the one prerequisite is, you need labeled data with the right sentiment categories.

Classification is a supervised learning technique, which means there is a lot of pre-processing involved in actually getting the labeled data, and that can be hard to set up. But once you have the labeled data, machine learning techniques are very scalable and can be very accurate in their prediction. This is a high-level overview of the training phase of a classification model. You have the input text with the right tags or labels, positive, neutral, negative. You tokenize, lemmatize, and extract features from the input text. You convert the input text to vector encodings or word embeddings. These are the features that you use to train the machine learning algorithm. Once you have a trained machine learning model, you can use it for prediction. Your prediction text will go through the same pre-processing techniques that you use for your training data. So you’ll tokenize it, lemmatize it, extract features, convert the words to vector representations or word embeddings, and then you’ll feed that into the model to get a prediction from the model.

Approaches to Sentiment Analysis

• Hybrid

Since both rule-based approaches and ML-based approaches have their own advantages, the hybrid approach seeks to combine both of them. It combines both of the approaches to get the best of both worlds. It uses features from both methods to optimize the speed and accuracy when deriving the contextual intent in text.

In practice, you may find that the hybrid approach, overall, gives you more accurate analysis than standalone approaches. However, it takes more time and effort to bring rule-based and machine learning-based systems together in a coherent way, so it may be harder to implement.

Word Vector Embeddings

Machine learning only works on numeric input. How do we represent text in numeric form?

{Refresh this page to see the gif video}

When you think about machine learning models, you know that ML models only process numeric data, and this is true of neural networks as well.

So the big question when you’re working with neural networks for sentiment analysis is how do we represent text in numeric form? Another way to put it, how do we express words as numbers so that ML models can understand them?

Let’s consider a document or text that we want to feed into a machine learning model for sentiment analysis. Here is a review of a restaurant. “The restaurant we visited last night was great, great food very good ambience.” Now the first thing you do is reprocess the document.. So all of the stop words are removed and the document is entirely in a lower case. You might get rid of all of the punctuation as well. You then tokenize this document where you’d extract tokens, which could be entire words or portions of words.

So let’s for simplicity, assume that every word in this document becomes a separate token after tokenization. Once you have these individual tokens from the document, the next step is for you to figure out how to represent these tokens numerically. You now need to represent each token in some numeric form.

So every bit of text or every document that you feed into your machine learning model is represented using a tensor of tokens where a tensor is just a multidimensional array.

Types of Sentiment Analysis

• Graded

• Aspect-based

• Intent-based

There are different types of sentiment analysis that you can perform on data and we’ll discuss a few of those types. There is graded sentiment analysis, aspect-based sentiment analysis, intent-based sentiment analysis, and emotional detection.

Let’s start with the first of these categories, graded sentiment analysis, and understand how this works. Graded sentiment analysis categorizes sentiment into multiple emotion levels. For example, you might think of sentiment as positive, neutral, negative, very negative or very positive, and you might grade these on a scale one to hundred. Graded sentiment analysis is commonly used in e-commerce systems via a five-star rating system to assess customer experience with products or services. This allows you to granularly express what you think of the product or service. This kind of sentiment analysis enables precision, engaging the polarity of the emotion. It’s ideal for businesses requiring detailed sentiment categorization, translating star ratings into corresponding sentiment levels. Having grades of emotion helps in understanding nuanced customer feedback by distinguishing between different degrees of positive and negative sentiment.

Next, we have aspect-based sentiment analysis. Now, if you are offering a product, there are different aspects or features of the product that customers might comment on. This kind of analysis concentrates on particular features or aspects of a product or service, like a laptop’s sound or keyboard quality. It tries to link customer sentiments to specific product attributes, identifying how each aspect is perceived. For example, if you buy a new laptop, you might like the look and feel of the laptop, but you might feel that the battery is not powerful enough. This is aspect-based sentiment analysis; figuring out what exactly you like and what you don’t for the same product. This kind of analysis assesses sentiments about individual features, distinguishing overall product sentiment from specific aspect sentiment. This provides detailed feedback for targeted improvements in different aspects of a product or service.

Next, we have intent-based sentiment analysis. This kind of sentiment analysis is widely used as a market research tool. This is utilized in understanding customer positions in the purchase cycle; very valuable for market research. This kind of sentiment analysis helps in crafting targeted marketing campaigns by picking up on specific customer interests, like discounts or deals on specific products. The idea to do intent-based sentiment analysis is to analyze sentiments to understand the likelihood of customer purchases, and it enables businesses to engage with customers based on their purchase readiness and interests.

Types of Sentiment Analysis

• Emotional detection

And finally, the fourth and last kind of sentiment analysis we’ll discuss today is emotional detection. This involves psychological state analysis. It goes beyond basic sentiment categories to analyze the writer’s psychological state. It tries to identify emotions on a broad emotional spectrum, identifies a range of emotions such as joy, anger, sadness, and frustration from the input text. It employs wordless or machine learning algorithms to discern the emotions conveyed by specific words. This analysis recognizes the diverse ways emotions can be expressed, accounting for context and varied expressions of feelings.

Challenges in Sentiment Analysis

Let’s discuss some challenges that you’d encounter in sentiment analysis.

It’s extremely difficult for machines to analyze sentiment in sentences that have sarcasm. Let’s say you say something like, “Yes, great, it took three weeks for my order to arrive.” Unless the sentiment analysis system has a deeper understanding of the context, it’s likely to label this as a positive sentence when it really isn’t.

Sentiment analysis systems also find it hard to deal with negation in sentences. This is the use of negative words to convey a reversal of meaning in the sentence. Let’s say you say something like, “I wouldn’t say this was a bad meal.” Sentiment analysis algorithms might have difficulty interpreting such sentences correctly.

Systems also find it hard to deal with multipolarity. This occurs when a sentence contains more than one sentiment. If a product review is something like, “I’m happy with the color of the product, but not happy with how long it took to deliver.” In order to extract the true sentiment of this sentence, you’ll need to use aspect-based sentiment analysis.

Sentiment analysis systems also find it hard to deal with comparisons. If you have text such as, “This product is second to none,” or, “This is better than nothing,” this is hard for systems to interpret.

Another challenge with sentiment analysis is defining what is meant by neutral. In all classification problems, defining your categories, and in this case, the neutral tag is one of the most important parts of the problem. What you mean by neutral, positive, or negative does matter when you train sentiment analysis problems. Objective texts might be considered neutral. Irrelevant information might be neutral. Text that contain wishes like, “I wish this was offered in a different color,” that might be neutral, but, “I wish this worked better, might not be neutral.