Sentiment analysis


Sentiment analysis is the use of natural language processing, text analysis, computational linguistics, and biometrics to systematically identify, extract, quantify, and study affective states and subjective information. Sentiment analysis is widely applied to voice of the customer materials such as reviews and survey responses, online and social media, and healthcare materials for applications that range from marketing to customer service to clinical medicine. With the rise of deep language models, such as RoBERTa, also more difficult data domains can be analyzed, e.g., news texts where authors typically express their opinion/sentiment less explicitly.

Types

A basic task in sentiment analysis is classifying the polarity of a given text at the document, sentence, or feature/aspect level—whether the expressed opinion in a document, a sentence or an entity feature/aspect is positive, negative, or neutral. Advanced, "beyond polarity" sentiment classification looks, for instance, at emotional states such as enjoyment, anger, disgust, sadness, fear, and surprise.
Precursors to sentimental analysis include the General Inquirer, which provided hints toward quantifying patterns in text and, separately, psychological research that examined a person's psychological state based on analysis of their verbal behavior.
Subsequently, the method described in a patent by Volcani and Fogel, looked specifically at sentiment and identified individual words and phrases in text with respect to different emotional scales. A current system based on their work, called EffectCheck, presents synonyms that can be used to increase or decrease the level of evoked emotion in each scale.
Many other subsequent efforts were less sophisticated, using a mere polar view of sentiment, from positive to negative, such as work by Turney, and Pang who applied different methods for detecting the polarity of product reviews and movie reviews respectively. This work is at the document level. One can also classify a document's polarity on a multi-way scale, which was attempted by Pang and Snyder among others: Pang and Lee expanded the basic task of classifying a movie review as either positive or negative to predict star ratings on either a 3- or a 4-star scale, while Snyder performed an in-depth analysis of restaurant reviews, predicting ratings for various aspects of the given restaurant, such as the food and atmosphere.
First steps to bringing together various approaches—learning, lexical, knowledge-based, etc.—were taken in the 2004 AAAI Spring Symposium where linguists, computer scientists, and other interested researchers first aligned interests and proposed shared tasks and benchmark data sets for the systematic computational research on affect, appeal, subjectivity, and sentiment in text.
Even though in most statistical classification methods, the neutral class is ignored under the assumption that neutral texts lie near the boundary of the binary classifier, several researchers suggest that, as in every polarity problem, three categories must be identified. Moreover, it can be proven that specific classifiers such as the Max Entropy and SVMs can benefit from the introduction of a neutral class and improve the overall accuracy of the classification. There are in principle two ways for operating with a neutral class. Either, the algorithm proceeds by first identifying the neutral language, filtering it out and then assessing the rest in terms of positive and negative sentiments, or it builds a three-way classification in one step. This second approach often involves estimating a probability distribution over all categories. Whether and how to use a neutral class depends on the nature of the data: if the data is clearly clustered into neutral, negative and positive language, it makes sense to filter the neutral language out and focus on the polarity between positive and negative sentiments. If, in contrast, the data are mostly neutral with small deviations towards positive and negative affect, this strategy would make it harder to clearly distinguish between the two poles.
A different method for determining sentiment is the use of a scaling system whereby words commonly associated with having a negative, neutral, or positive sentiment are given an associated number on a −10 to +10 scale or simply from 0 to a positive upper limit such as +4. This makes it possible to adjust the sentiment of a given term relative to its environment. When a piece of unstructured text is analyzed using natural language processing, each concept in the specified environment is given a score based on the way sentiment words relate to the concept and its associated score. This allows movement to a more sophisticated understanding of sentiment, because it is now possible to adjust the sentiment value of a concept relative to modifications that may surround it. Words, for example, that intensify, relax or negate the sentiment expressed by the concept can affect its score. Alternatively, texts can be given a positive and negative sentiment strength score if the goal is to determine the sentiment in a text rather than the overall polarity and strength of the text.
There are various other types of sentiment analysis, such as aspect-based sentiment analysis, grading sentiment analysis, multilingual sentiment analysis and detection of emotions.

Subjectivity/objectivity identification

This task is commonly defined as classifying a given text into one of two classes: objective or subjective. This problem can sometimes be more difficult than polarity classification. The subjectivity of words and phrases may depend on their context and an objective document may contain subjective sentences. Moreover, as mentioned by Su, results are largely dependent on the definition of subjectivity used when annotating texts. However, Pang showed that removing objective sentences from a document before classifying its polarity helped improve performance.
The term objective refers to the incident carrying factual information.
The term subjective describes the incident contains non-factual information in various forms, such as personal opinions, judgment, and predictions, also known as 'private states'. In the example down below, it reflects a private states 'We Americans'. Moreover, the target entity commented by the opinions can take several forms from tangible product to intangible topic matters stated in Liu. Furthermore, three types of attitudes were observed by Liu positive opinions, 2) neutral opinions, and 3) negative opinions.
  • Example of a subjective sentence: 'We Americans need to elect a president who is mature and who is able to make wise decisions.'
This analysis is a classification problem.
Each class's collections of words or phrase indicators are defined for to locate desirable patterns on unannotated text. For subjective expression, a different word list has been created. Lists of subjective indicators in words or phrases have been developed by multiple researchers in the linguist and natural language processing field states in Riloff et al.. A dictionary of extraction rules has to be created for measuring given expressions. Over the years, in subjective detection, the features extraction progression from curating features by hand to automated features learning. At the moment, automated learning methods can further separate into supervised and unsupervised machine learning. Patterns extraction with machine learning process annotated and unannotated text have been explored extensively by academic researchers.
However, researchers recognized several challenges in developing fixed sets of rules for expressions respectably. Much of the challenges in rule development stems from the nature of textual information. Six challenges have been recognized by several researchers: 1) metaphorical expressions, 2) discrepancies in writings, 3) context-sensitive, 4) represented words with fewer usages, 5) time-sensitive, and 6) ever-growing volume.
  1. Metaphorical expressions. The text contains metaphoric expression may impact on the performance on the extraction. Besides, metaphors take in different forms, which may have been contributed to the increase in detection.
  2. Discrepancies in writings. For the text obtained from the Internet, the discrepancies in the writing style of targeted text data involve distinct writing genres and styles.
  3. Context-sensitive. Classification may vary based on the subjectiveness or objectiveness of previous and following sentences.
  4. Time-sensitive attribute. The task is challenged by some textual data's time-sensitive attribute. If a group of researchers wants to confirm a piece of fact in the news, they need a longer time for cross-validation, than the news becomes outdated.
  5. Cue words with fewer usages.
  6. Ever-growing volume. The task is also challenged by the sheer volume of textual data. The textual data's ever-growing nature makes the task overwhelmingly difficult for the researchers to complete the task on time.
Previously, the research mainly focused on document level classification. However, classifying a document level suffers less accuracy, as an article may have diverse types of expressions involved. Researching evidence suggests a set of news articles that are expected to dominate by the objective expression, whereas the results show that it consisted of over 40% of subjective expression.
To overcome those challenges, researchers conclude that classifier efficacy depends on the precisions of patterns learner. And the learner feeds with large volumes of annotated training data outperformed those trained on less comprehensive subjective features. However, one of the main obstacles to executing this type of work is to generate a big dataset of annotated sentences manually. The manual annotation method has been less favored than automatic learning for three reasons:
  1. Variations in comprehensions. In the manual annotation task, disagreement of whether one instance is subjective or objective may occur among annotators because of languages' ambiguity.
  2. Human errors. Manual annotation task is a meticulous assignment, it require intense concentration to finish.
  3. Time-consuming. Manual annotation task is an assiduous work. Riloff show that a 160 texts cost 8 hours for one annotator to finish.
All these mentioned reasons can impact on the efficiency and effectiveness of subjective and objective classification. Accordingly, two bootstrapping methods were designed to learning linguistic patterns from unannotated text data. Both methods are starting with a handful of seed words and unannotated textual data.
  1. Meta-Bootstrapping by Riloff and Jones in 1999. Level One: Generate extraction patterns based on the pre-defined rules and the extracted patterns by the number of seed words each pattern holds. Level Two: Top 5 words will be marked and add to the dictionary. Repeat.
  2. Basilisk by Thelen and Riloff. Step One: Generate extraction patterns. Step Two: Move best patterns from Pattern Pool to Candidate Word Pool. Step Three: Top 10 words will be marked and add to the dictionary. Repeat.
Overall, these algorithms highlight the need for automatic pattern recognition and extraction in subjective and objective task.
Subjective and object classifier can enhance the several applications of natural language processing. One of the classifier's primary benefits is that it popularized the practice of data-driven decision-making processes in various industries. According to Liu, the applications of subjective and objective identification have been implemented in business, advertising, sports, and social science.
  • Online review classification: In the business industry, the classifier helps the company better understand the feedbacks on product and reasonings behind the reviews.
  • Stock price prediction: In the finance industry, the classifier aids the prediction model by process auxiliary information from social media and other textual information from the Internet. Previous studies on Japanese stock price conducted by Dong et al. indicates that model with subjective and objective module may perform better than those without this part.
  • Social media analysis.
  • Students' feedback classification.
  • Document summarising: The classifier can extract target-specified comments and gathering opinions made by one particular entity.
  • Complex question answering. The classifier can dissect the complex questions by classing the language subject or objective and focused target. In the research Yu et al., the researcher developed a sentence and document level clustered that identity opinion pieces.
  • Domain-specific applications.
  • Email analysis: The subjective and objective classifier detects spam by tracing language patterns with target words.