Exploration of Competitive Market Behavior Using Near-Real-Time Sentiment Analysis

Contents

I. Abstract

II. Acknowledgements

III. List of Figures

IV. List of Tables

V. List of Equations

VI. Abbreviations

1. Introduction
1.1. Goal of the Thesis
1.2. Chapter Outline

2. Sentiment Analysis
2.1. Goals of Sentiment Analysis
2.2. Use Cases of Sentiment Analysis
2.3. Challenges of Processing Natural Language
2.3.1. Negation
2.3.2. Deontic Irrealis
2.3.3. Languages
2.3.4. Emoticons, Acronyms, and Further Improvements
2.4. Domain-specific Language
2.5. Algorithmic Principles of Sentimental Analysis
2.6. Measuring Sentiment Analysis Accuracy
2.6.1. Precision, Recall and Accuracy
2.6.2. F-score
2.6.3. Supervised Machine-Learning Algorithms
2.6.4. Unsupervised Algorithms
2.6.5. Dictionary-based Algorithms
2.6.6. Comparison of the Sentiment Analysis Algorithms
2.7. Sentiment Analysis in Social Networks

3. Large-Scale Sentiment Analysis System
3.1. System Architecture Overview
3.2. Data Scraping
3.2.1. Data Structure
3.2.2. Twitter’s Streaming API
3.3. Scalability with Hadoop
3.3.1. Scalable Data Storage: HDFS and HBase
3.3.2. Scalable Computation: MapReduce and Derivatives
3.4. Sentiment Analysis Implementation
3.4.1. Overview of Data Computing Process
3.4.2. Pre-Processing: Cleaning, Tokenization and POS-Tagging
3.4.3. Processing: Computing the Basic Sentiment Score
3.4.4. Post-Processing: Enhancing Sentiment Analysis Score Precision
3.4.5. Scheduling and Modularity for Near-Real-Time Predictions
3.5. Evaluation of Sentiment Analysis Accuracy
3.6. Evaluation of Speed and Scalability
3.7. Providing Results – Using the Calculated Sentiments
3.7.1. Aggregated Sentiment Values and Near-Real-Time Analysis
3.7.2. Preparing for Basic Online Analytical Processing

4. Commercial Use Cases: Attaining Competitive Advantages
4.1. Business Brand Value Monitoring
4.1.1. Near-Real-Time Monitoring
4.1.2. Historical Analysis
4.1.3. Demographic and Geolocation-based Analyses
4.2. Sentiments of Brands, Products and Markets
4.3. Sentiments of Product Features

5. Scientific Use Cases: Understanding Competitive Market Behavior
5.1. Fundamentals of Economic Markets
5.1.1. Market Forms
5.1.2. Market Maturity
5.2. Sentiments in Competitive Markets
5.2.1. First Hypothesis
5.2.2. Second Hypothesis
5.2.3. Third Hypothesis

6. Conclusion

6.1. Limitations of Findings and Further Work

7. Bibliography

I. Abstract

Emotions are not only the foundation of human life, but also influence all decisions in modern markets. Grasping the reasoning behind our choices is a key element in economic sciences. Sentiment analysis, a tool to extract emotions from text, is used in this thesis to analyze customers’ opinions in various markets. The calculations are done on a server architecture that is designed to be scalable for massive input directly from social networks. It computes the sentiment score in a flexible multi-stage process and provides several methods of accessing the results.

Subsequently, it is demonstrated how to use the system’s capabilities by implementing various commercial use cases. This includes geographical and demographic analysis. Additionally, the system is able to provide near-real-time results.

Lastly, the thesis concludes by performing several correlation analyses on the collected data. This illustrates how the intensity of emotions vary by the maturity and form of the economic market and affects the participating companies in these markets.

II. Acknowledgements

First of all, I would express my gratitude for all the assistance I received from Prof. Dr. Reiner Dumke. Not only did he advise my bachelor thesis and master thesis, he also provided endless help in my endeavors for two adventures abroad. Furthermore, Prof. Dumke was also a reason that I received a full scholarship.

Dr. Robert Neumann guided and advised me during the course of my advancement through academia, which now finalizes in this master thesis. I feel very thankful for the inspiration, which not only steered me through this thesis, but also through the last few years.

This master thesis was primarily researched and written while I was living in Stevens Point, USA. My internship at the local university was supervised by Scott Gile, who did more than manage my work. He also encouraged, inspired, and mentored me on a daily basis. I feel extraordinarily grateful for the experience I gained that year.

Another inspiration of the last year in Wisconsin was my constant motivation from Jesse Kosobucki. Thank you for inciting me to continue to work on myself and being extremely patient through my long working hours on this thesis.

None of that would have been possible without the continuous support from my parents for the past 25 years. All of the experiences of the past, which eventually lead to this master thesis, were permitted by the infinite help I received.

III. List of Figures

Figure 1 – Designed Process for a Large-Scale Sentiment Analysis System

Figure 2 - Twitter's Streaming API [66]

Figure 3 – Composition and Basic Principle of MapReduce

Figure 4 – Server Architecture of Scalable Sentiment Analysis System

Figure 5 – Overview of Sentiment Scoring Process

Figure 6 – Converting Single Tweet to Aggregated Sentiment Score

Figure 7 – Decrease in Computation Time with More VMs

Figure 8 – Potential and Use Cases of the Sentiment Analysis System

Figure 9 – Visualization of Sentiment Score and Volume

Figure 10 – Historical View of Sentiment Score and Tweet Count

Figure 11 – Pie Chart of Tweets by Sentiment Score (Monthly Summary)

Figure 12 – Sentiment Score Executive Report

Figure 13 – Heat Map of Tweet Count for a Business in Melbourne

Figure 14 – Average Sentiment Score (Monthly) for Multiple Products

Figure 15 – Pie Chart of Relevant Product Features for Sentiment Category

IV. List of Tables

Table 1 – Emoticons and Their Meanings

Table 2 – Acronyms and the Corresponding Words

Table 3 – Confusion Table

Table 4 – Demo for Calculating Sentiment Scores

Table 5 – Result Comparison between Sentiment Algorithms [53]

Table 6 – Search Keywords for Smartphone Manufacturer Market

Table 7 – Search Keywords for Computer Operating Systems

Table 8 – Search Keywords for Premium German Car Manufacturer

Table 9 – Evaluation of Implemented Sentiment Scoring Algorithm

Table 10 – Pre-Processing and Processing: Computation Times

Table 11 – Post-Processing: Computation Times

Table 12 – Decrease in Computation Time with Added VMs

Table 13 – Database Design for Aggregated Sentiment Values

Table 14 – Sentiment Scores by Region and Product

Table 15 – Classification of the Analyzed Markets

Table 16 – Absolute Sentiment Score – Correlation between Audi and BWM

Table 17 – Absolute Sentiment Score – Correlation between Mercedes and BWM

Table 18 – Absolute Sentiment Score – Correlation between Audi and Mercedes

Table 19 – Absolute Sentiment Score – Correlation between Windows and Linux

Table 20 – Absolute Sentiment Score – Correlation between Windows and Mac OS X

Table 21 – Absolute Sentiment Score – Correlation between Mac OS X and Linux

Table 22 – Absolute Sentiment Score – Correlation between Android and Windows Phone

Table 23 – Absolute Sentiment Score – Correlation between iOS and Windows Phone

Table 24 – Absolute Sentiment Score – Correlation between iOS and Android

Table 25 – Relative Sentiment Score – Correlation between BMW and Audi

Table 26 – Relative Sentiment Score – Correlation between BMW and Mercedes

Table 27 – Relative Sentiment Score – Correlation between Mercedes and Audi

Table 28 – Relative Sentiment Score – Correlation between Windows and Linux

Table 29 – Relative Sentiment Score – Correlation between Windows and Mac OS X

Table 30 – Relative Sentiment Score – Correlation between Linux and Mac OS X

Table 31 – Relative Sentiment Score – Correlation between Android and Windows Phone

Table 32 – Relative Sentiment Score – Correlation between iOS and Windows Phone

Table 33 – Relative Sentiment Score – Correlation between Android and iOS

Table 34 – Relative Sentiment Score – Correlation between Arsenal and Chelsea F.C.

Table 35 – Correlation between Android and iOS (Top 25 Events)

Table 36 – Correlation between Audi and BMW (Top 25 Events)

Table 37 – Correlation between Audi and Mercedes (Top 25 Events)

Table 38 – Correlation between BMW and Mercedes (Top 25 Events)

Table 39 – Indications for Correlation Strengths based on Market Types

V. List of Equations

Equation 1 – Computation of Precision

Equation 2 – Computation of Recall

Equation 3 – Computation of Accuracy

Equation 4 – Computation of F-score [45]

Equation 5 – Computation of F-score with Confusion Table

Equation 6 – Balanced F-score

Equation 7 – Sentiment Score Objectivity

Equation 8 – Sentiment Score Polarity

Equation 9 – Tweet Aggregation Scale: Subjectivity

Equation 10 – Tweet Aggregation Scale: Polarity

VI. Abbreviations

Abbildung in dieser Leseprobe nicht enthalten

1. Introduction

Humanity has come a long way in the last few thousand years. One fact stayed true throughout time: our opinions are the fundamentals for all of our activities. Our beliefs, our perceptions of the world around us, and the emotions processing that information are a key factor which set us humans apart from other living creatures. Our state of mind has and permanently will shape our lives on a day-to-day basis. Although it is such a central part of our existence, we have a hard time grasping the concept. We design and label discrete categories of emotions and feelings to better explain ourselves. However, the accuracy is thoroughly limited. If someone says he is happy, how happy is he? Why is he happy? Humans have a hard time quantifying and reasoning their feelings.

In the modern world, this becomes especially clear in economic markets. Our emotions are often the deciding factor of our actions in marketplaces, even though we usually pretend to argue with the facts of the product features. The decisions for—or against—a product or service are often ultimately made by our subconscious viewpoints. Understanding that behavior, and the partially hidden reasoning, would give us the opportunity to further comprehend ourselves and our actions better.

Unfortunately, it is virtually impossible to track our emotions during our day-to-day life. Luckily, we have developed languages to express ourselves. The work in this thesis endeavors to use that indirect link to our state of mind to quantify our feelings. Extracting human emotions and feelings from language, ideally text, is called sentiment analysis [1]. Later, the collected and processed data attempts to grasp our sentiments in economic markets. Overall, the work in this thesis ties multiple areas of research together: it requires computer science to design a capable server architecture and it uses knowledge from economics and business theory to analyze the results of that system. Finally, it can be further expanded into areas of consumer behavior and social sciences.

1.1. Goal of the Thesis

The goal of this thesis is to use publically available text data of consumers to understand their sentiments and behaviors in a market place. In order to be capable of evaluating customers’ sentiments, a system needs to be implemented to perform natural language processing tools to extract the sentiment from a written text. Another goal of the system is to be extremely fast to achieve near-real-time results¹. Furthermore, the system also needs to handle large amount of data. While none of the said techniques are revolutionary, the tools were never tied together into one system to achieve something unique. The complex development is structured in three technical phases, while two supplementary analysis phases exploit the system for commercial and scientific interests.

Initially, in phase one, it requires the development of a sentiment analysis system, which is close to state-of-the-art technology. The algorithm should be able to analyze and predict the sentiment of a written sentence, paragraph, or of an entire document. It utilizes a scalable algorithm, which can process substantial amounts of data.

The second phase includes the utilization of this algorithm in a large-scale sentiment analysis system. It adds components to the sentiment analysis to store the raw texts in a data storage with high-read performance. It also is able to analyze a continuous stream of incoming texts. The system should be optimized towards a source with a promising potential for a constant stream of generated user data, e.g. a social network.

The third phase is to prepare the system for further usage. This can be commercial or scientific interests. This covers the addition of high performance database querying for analytics. It also includes offering real-time monitoring of sentiment values. For the latter, the system needs to offer easy access and visual tools for exploration of variations in sentiments.

Finally, in phase four, the data of the system is used to analyze current markets and answer several customer behavioral questions. This includes attempting to understand the actions and sentiments of consumers regarding competing brands.

1.2. Chapter Outline

The master thesis is structured in several chapters. The first chapter Sentiment Analysis will cover the fundamental objectives, principles, and challenges of sentiment analysis. Before including several algorithmic methods of understanding sentiments of written text, it introduces a number of measures to compare accuracy. The argument for one process to be implemented in phase one will be made.

The next chapter, Large-Scale Sentiment Analysis System, covers development phases two and three. It explains how to access streams of social network data, store, and analyze them in a scalable fashion. The chapter also reasons for the choice of a specific structure to handle the scalability challenge. It additionally illustrates the steps of a high performance sentiment analysis process. In order to compare the functioning of the system, different evaluations will be made. Lastly, the system is prepared for scientific and commercial uses.

The second functionality will be exploited in chapter Commercial Use Cases: Attaining Competitive Advantages. This section implements a structured view of the data to gain new knowledge about customers and their opinions. It will also apply the near-real-time capabilities of the system to reveal current developments in economic markets.

Lastly, the system is used to attempt to gain information about consumer behavior in market situations. Several questions are articulated and the sentiment data, including the query interfaces, are used to attempt to reason for the consumers’ behavior.

2. Sentiment Analysis

With the rise of information technology, the amount of documents available on digital storage is vastly increasing. While shifting from traditional paper documents to electronic data, the way humans organize their files is changing. Old-fashioned processes for filing, sorting, and finding paper documents cannot be applied to their electronic counterparts. Electronic filing has the advantage that there are numerous ways of indexing and discovering the data.

While the possibilities increased, at least at the beginning of electronic filing, individuals often sorted their documents by topic. Subsequently, since the establishment of electronic files, researchers have used machine learning techniques to automatically classify documents into categories. That was the beginning of algorithms which go through natural text, analyze, and assign a numerical or text value depending on the content. Over the years, the use cases and goals varied, but the basic idea—algorithmically understanding the content of written natural language—stayed the same. Initially, numerically quantifying the sentiment of a text was not part of the goal.

However, the last two decades saw an enormous increase of Internet use and the widespread of websites with user generated content. Content based on reviews of products, movies, or books has especially gained significant popularity. In 2002, Pang et al. ² suggested to transfer the topic classification methods to review websites in order to automatically classify a review as positive or negative. The main goal was to produce a label or short summary of the text. Similarly and also in 2002, Turney ³ automatically applied recommended or not recommended to reviews. These were the first very basic applications of sentiment analysis.

Sentiment analysis, sometimes also called opinion mining, refers to attempting to determine the attitude and emotions underlying the text written by a human being. The term describing the research area changed over time and even today, these still are interchangeably used terms for it ⁴. This thesis will use sentiment analysis.

In 2002, Turney [2] demonstrated that standard machine learning techniques can outperform human-produced baselines of understanding human emotions in movie reviews by classifying documents into a positive or negative sentiment. The authors applied a variety of different machine learning algorithms. While they have small differences in performance, they all outperform proposed classifications by humans. The overall accuracy rate for the human baseline (50% to 69%) is low [2]. This reveals the complexity of understanding natural language.

In 2004, Pang et al. ⁵ presented one of the first complex algorithms that “seeks to identify the viewpoint(s) underlying a text span.” Again, the domain was movie reviews. For early research, the domain of movie reviews was ideal because the website requested a thumbs up or a thumbs down as a summary for the text. The classifier then attempted to predict the result by analyzing the written review.

2.1. Goals of Sentiment Analysis

Sentiment analysis goes beyond the previous examples which associate a short text with a semantic direction, i.e. either negative or positive. The classification to the negative/positive scale is a strong simplification of human emotions, which are very complex. Connecting a text to a most accurate emotion is a challenging task. The latest research tries to specify the intensity of the emotion, for example weak positive and strong positive [4].

Another goal is to understand whether the text contains polarity or not. For example, stating a fact does not contain any emotion, while announcing personal feelings unfolds the writer’s emotional state. Nevertheless, emotions go beyond a positive/negative or neutral/emotional scale.

Whitelaw et al. ⁶ researched appraisal groups in context of sentiment analysis. The researchers additionally added the attitude (affect, appreciation or judgment) and graduation (force or focus) to orientation (positive/negative) and polarity (marked or unmarked). The combination of the four attributes allowed the research team to further understand the emotional state.

The previous goals just predict the sentiment. A further goal of sentiment analysis is to understand why a certain text is positive or negative. For example, a review of a smartphone could be moderately positive, because of the excellent size and battery life, but not perfect because of the low quality of the camera. Understanding contextual aspects is a difficult task, which can be very valuable for commercial applications ⁷.

Another area of concentration is the comparison between two entities ⁸. The sentiment analysis algorithm attempts to quantity which product the user prefers based on the given text. Lastly, the last two aspects can be combined. For example, Camera A is better than Camera B, because the user likes the lighter weight, which is important to him.

Emotions are a large element of human life, but are also a not even remotely comprehensible component. Humans sometimes struggle recognizing their own emotions. Understanding the thought processes and feelings of others is not always mutually intelligible. The idea of having a neutral evaluator of an emotion, a computer algorithm, is a thought-provoking idea.

The goals of sentiment analysis developed over the years. In the first years, it was a challenge for the scientific community even though ⁹ predicted in 2003 that sentiment analysis has the potential for a “competitive analysis, marketing analysis, and detection of unfavorable rumors for risk management.” The technology was under heavy research, but companies did not have easy access to the required data. The further increase of fast Internet access and the growing popularity of online communities started to provide a large pool of data to tap into. Product review websites provided a range of negative and positive texts. However, these were biased and a selection of human emotions. Finally, the rise of social networks offers an almost unfiltered stream of human emotions in written form.

2.2. Use Cases of Sentiment Analysis

The first applications were using sentiment analysis to classify reviews on websites and text documents. This basically can be done on any topic or domain ¹⁰. Popular examples are hotels ¹¹, movies [2], ¹², politics ¹³, product reviews ¹⁴, and news ¹⁵, ¹⁶. Often the analysis is done in a two dimensional way—the sentiment is either positive or negative. Sometimes the polarity scale (emotional or factual) is also analyzed.

In the following years, sentiment analysis developed further uses, especially for businesses. Companies had a strong interest in measuring the user’s feelings towards their products. Traditionally, this had to be done in costly surveys and it was complicated to communicate to unsatisfied individual costumers. When the Internet started to be a platform for exchanging opinions and experiences with products, a large opportunity emerged.

In the past, a complaining customer was hard to detect through classic polling and impossible to mitigate. Nowadays, if a customer complains on the Internet, the company of the product can contact the person and offer help or compensation. This can significantly improve the satisfaction and word-of-mouth.

If sentiment analysis is used to automatically classify large quantities of feedback and reviews, the company can develop many measurements of customer satisfaction. It gives them the opportunity to closely monitor and examine the sentiment of their customers. In the long run, this can be essential for any company.

Another approach for the company is to understand themes of products and markets. This can be done on multiples dimensions, not just positive or negative ¹⁷. Summarization of reviews is another element of sentiment analysis ^{18 19}. Monitoring the market also includes understanding success factors of the products from the competition.

The recent years exhibited the importance of online review platforms for users’ purchase decisions. Negative reviews can significantly influence a company’s success in a market. Competitors are aware of the potential to impact the market by adding fake reviews on websites. Jindal and Liu ²⁰ researched and showed the wide-spread of “opinion spam.” They also identify different kinds of spam and propose detection algorithms to countervail the effects of opinion spam.

Sentiment analysis does not have to be within a business background. Politicians, especially during election periods, are closely monitoring their success on social networks. Monitoring the change over time gives them a tool to adjust their strategies. The opportunity also attracted election prediction researcher. Tumasjan et al. ²¹ used Twitter as a reflection of the users’ political sentiment to correctly predict the outcome of the 2010 German Election

Sentiment analysis is also a very interesting field for market research. Similar to a business, the research team can replace or complement classic polling strategies to collect data about customers’ opinions. Huberman and Asur ²² shows that it is possible to predict the box revenue of movies just by listening to statuses on the social network Twitter. It requires analyzing the pre-release attention and release-phase attention to calculate the success of a movie. The research concluded that the prediction based on sentiment analysis in social networks can outperform classic polling in accuracy, time effectiveness, and costs. Another scientific discovery is that Twitter even is able to predict the stock market ²³.

While the opportunities are huge, there are risks involved. An excellent example of sentiment analysis with undesired side effects is the Hathaway case. As earlier mentioned, the mood on Twitter can predict the stock market. Anne Hathaway, an actress, and Berkshire Hathaway, a company, might be a perfect show case of such a side effect. The data suggests, that whenever Anne Hathaway is in a positive way in the news, the Berkshire company stocks went up ²⁴. This is a correlation which could be caused by automatic trading software on the stock market, which indicates that the software takes sentiments on social networks into account.

The various use cases demonstrate the need for a precise calculation of users’ sentiment. There are a few challenges to calculate the sentiment value of a written text. The next chapter will introduce several aspects of the complexity of processing natural language.

2.3. Challenges of Processing Natural Language

As described in an earlier section, sentiment analysis attempts to fulfill a variety of use cases by achieving several aspects of recognizing sentiment in human language. This is a very difficult task. Language is very imprecise and contains several aspects which can strongly change or even reverse the meaning of an entire paragraph. Humans use these linguistic devices to express themselves more strongly. Additionally, written text is a very noisy area ²⁵. An entire article can state hundreds of sentences which just objectively state facts and do not have any subjective sentiment and then one sentence can change the entire, overall sentiment.

The sentiment analysis algorithm needs to differentiate between facts and opinions, and also understand how strong these opinions are. However, emotions do not have to be stated directly. Humans can express themselves by using negation, irrealis, or sarcasm. This chapter will explore several features of human language. It explains why it is difficult to detect these aspects and possibly how to mitigate troublesome characteristics of human language.

2.3.1. Negation

One of the most important linguistic devices is negation. Language negators (i.e. “not,” “without,” and “lacks”) reverse the polarity of a sentence and are a complicated detail of sentiment analysis. The following excerpt from a movie review ²⁶ demonstrates the power of negators:

This film should be brilliant. It sounds like a great plot, the actors are first grade, and the supporting cast is good as well, and Stallone is attempting to deliver a good performance. However, it can’t hold up.

The quote features several clear-cut positive words and a single imperceptible negator reverses the entire sentiment of the review. Negation is a very powerful tool in the human language, but extremely hard to detect and understand for algorithms ²⁷. Research agrees that they play an important role in the overall accuracy ²⁸, ²⁹.

However, the main goal of this master thesis is the analysis of overall sentiment value of a company’s brand in a specific time frame. The individual sentiment of a single text is less critical, since there is an aggregation of thousands of sentiment values into one score. Thus, the negation—in this specific use case—is of less impact, because it will overall offset when aggregating the sentiment values. Nonetheless, in the future it is necessary to add a negation analysis module.

Another variation of negation is sarcasm or irony. That is “classically defined as the rhetorical process of intentionally using words or expressions for uttering a meaning different (usually the opposite) from the one they have when used literally” ³⁰. Detecting irony is possible, if emoticons (see section Emoticons, Acronyms, and Further Improvements) or ratings are used which have the opposite meaning of the text ³¹.

2.3.2. Deontic Irrealis

Deontic irrealis is a linguistic term for describing a failed expectation or desire. This is particularly difficult to detect and score in the sentiment analysis process. For example:

“This movie should have been good.”

This sentence has “good” as an indicator for positive sentiment and no direct pointers for a negative sentiment [4]. However, the sentiment of the writer is towards the negative side, since his high expectations have not been met and he thus holds a negative sentiment of disappointment.

2.3.3. Languages

So far, one important factor regarding sentiment analysis was ignored—language. Throughout the world, there are a variety of languages that are represented on the Internet, which is data source of this master’s thesis. Depending on the used algorithm, the transfer to another language could be in different forms. The supervised machine-learning classifiers need to be trained on a new dataset. If the system uses a dictionary-based approach, an entire new dictionary needs to be created for the language. Parts of some algorithms are based on language features which are specific to Western languages.

Most research has been done on Western languages, especially English. These languages have high similarities in syntax and grammar. Asiatic languages, for example, have a significantly different syntax and require a vastly modified implementation of the sentiment analysis algorithm. The Chinese language, with vastly different language syntax and word constructs, has been researched successfully ³².

Ideally, the system will be multilingual by design. Denecke ³³ and Hiroshi et al. ³⁴ accomplished this by automatically translating reviews to English and then running sentiment analysis on the translated text. Bautin et al. ³⁵ used a similar approach to compare the different reactions to world news in nine societies. However, this system is limited by the quality of the machine translation and the amount of sentiment becoming lost in the translations.

Boyd-Graber and Resnik ³⁶ implemented a multilingual supervised latent Dirichlet allocation, which results in a "probabilistic generative model that allows insights gleaned from one language’s data to inform how the model captures properties of other languages.” This is the foundation to make sentiment analysis algorithms work across languages. Abase et al. ³⁷ implemented an actual system to work with English and Arabic forum posts. Even though the languages are fundamentally diverse, they reached a common accuracy over both languages of over 90%. However, human communication goes beyond the use of words, independent of the used language.

2.3.4. Emoticons, Acronyms, and Further Improvements

Generally, non-verbal expressions are lost when using written communication. Yet, humans can attempt to replace facial expressions and moods via emoticons. Emoticons are “graphic representations of facial expressions” ³⁸. An example selection sorted by frequency of use is viewable in Table 1 [39].

Table 1 – Emoticons and Their Meanings

Abbildung in dieser Leseprobe nicht enthalten

Unlike words, emoticons have low ambiguity in their sentimental meaning. Thus, they are an excellent opportunity to improve the classification precision of a sentiment algorithm. Past research proves that the consideration of emoticons in texts can improve the results ³⁹, ⁴⁰.

However, the use of emoticons is not the only change of language featured when humans communicate via written words. People have formed a whole new subset of terms. One element is the extensive use of acronyms, especially by replacing phonetic sounds by numbers with the identical phonetic sound. Table 2 provides examples. Aue and Gamon [40] reassigned acronyms into their original meaning and achieved positive results.

Table 2 – Acronyms and the Corresponding Words

Abbildung in dieser Leseprobe nicht enthalten

The scientific community did not stop at these obvious chances to further improve the accuracy. The human language offers many aspects to learn the meaning with a machine. An example is the understanding of target-specific adjectives by Fahrni and Klenner ⁴¹ or added adverbs by Benamara et al. ⁴².

2.4. Domain-specific Language

The previous sections were very general and based on language granularities. However, humans tend to use a different type of language depending on the domain. Even if the algorithms understand the words and syntax correctly, the result can vary between various topics. The majority of sentiment classification algorithms are supervised and directly or indirectly based on a dictionary—an association of word(s) to a positive or negative meaning. These dictionaries are often built to a specific domain to accomplish the highest possible outcome. If the text data source is shifted into a different domain, where another type of language is used, a significant drop in accuracy is observed ⁴³.

Taboada et al. [4] show that movies, which are based off of video games, are largely not well received. When a reviewer mentions the word “game,” it is often a strong negative indicator, but this only applies for movie reviews. In other domains, the word “game” is just noise without a particular sentiment.

When shifting the domain, the previously used training data could render partly or completely inoperable. Building the dictionary or a training set is the largest factor of effort when implementing a system to run sentiment analyses. If it is required to build a completely new dictionary for every domain, the effort multiplies. One approach is to use the training data from the previous domain and to transfer it to the new domain ⁴⁴. Another approach is to have one large training set instead of many smaller but separated ones. Aue and Gamon [40] investigated combining the dictionaries of the various fields—books, movies, and electronics—while maintaining the level of accuracy. The last chapter clarified how challenging sentiment analysis is. In order to be effective, the algorithm needs to be accurate and low on errors. The next section introduces several algorithms and approaches to conclude human emotions from text.

[...]

---

¹ B. Liu, Sentiment Analysis and Opinion Mining (Synthesis Lectures on Human Language Technologies), Morgan & Claypool Publishers, 2012.

² B. Pang, L. Lee and S. Vaithyanathan, "Thumbs up?: sentiment classification using machine learning techniques," in ACL-02 conference on Empirical methods in natural language processing, 2002.

³ P. D. Turney , "Thumbs up or thumbs down?: semantic orientation applied to unsupervised classification of reviews," in Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, Philadelphia, Pennsylvania, 2002.

⁴ M. Taboada, J. Brooke, M. Tofiloski, K. Voll and M. Stede, "Lexicon-based methods for sentiment analysis," Computational Linguistics, vol. 2, no. 267-307, p. 37, 2011.

⁵ B. Pang and L. Lee, "A sentimental education: sentiment analysis using subjectivity summarization based on minimum cuts," in Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics, Barcelona, Spain, 2004.

⁶ C. Whitelaw, N. Garg and S. Argamon , "Using appraisal groups for sentiment analysis," in 14th ACM international conference on Information and knowledge management, Bremen, Germany, 2005.

⁷ H. Shi, G. Zhou and P. Qian, "An Attribute-based Sentiment Analysis System," Information Technology Journal, vol. 9, pp. 1607-1614, 2010.

⁸ W. Kessler and J. Kuhn, "A Corpus of Comparisons in Product Reviews," in International Conference on Language Resources and Evaluation, Reykjavik, Iceland, 2014.

⁹ T. Nasukawa and J. Yi, "Sentiment analysis: capturing favorability using natural language processing," in Proceedings of the 2nd international conference on Knowledge capture, Sanibel Island, FL, USA, 2003.

¹⁰ P. Kherwa, A. Sachdeva, D. Mahajan, N. Pande and P. K. Singh, "An approach towards comprehensive sentimental data analysis and opinion mining," in IEEE International Advance Computing Conference (IACC), Gurgaon, 2014.

¹¹ W. Kasper and M. Vela, "Sentiment Analysis for Hotel Reviews," in Computational Linguistics-Applications Conference, 2011.

¹² A. Kennedy and D. Inkpen, "Sentiment classification of movie reviews using contextual valence shifters," in Computational Intelligence, 2006.

¹³ T. Mullen and R. Malouf, "A Preliminary Investigation into Sentiment Analysis of Informal Political Discourse," in AAAI Spring Symposium: Computational Approaches to Analyzing Weblogs, 2006.

¹⁴ A.-M. Popescu and O. Etzioni, "Extracting Product Features and Opinions from Reviews," in Natural language processing and text mining, London, Springer, 2007, pp. 9-28.

¹⁵ T. Weninger, X. A. Zhu and J. Han, "An exploration of discussion threads in social news sites: a case study of the Reddit community," in IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining, Niagara, Ontario, Canada, 2013.

¹⁶ N. Godbole, M. Srinivasaiah and S. Skiena, "Large-Scale Sentiment Analysis for News and Blogs," in ICWSM 7, 2007.

¹⁷ S. Thanangthanakij, E. Pacharawongsakda, N. Tongtep, P. Aimmanee and T. Theeramunkong, "An Empirical Study on Multi-dimensional Sentiment Analysis from User Service Reviews," in Seventh International Conference on Knowledge, Information and Creativity Support Systems (KICSS), Melbourne, VIC, 2012.

¹⁸ L. Zhuang, F. Jing and X.-Y. Zhu, "Movie review mining and summarization," in 15th ACM international conference on Information and knowledge management, Arlington, Virginia, USA, 2006.

¹⁹ M. Hu and B. Liu , "Mining and summarizing customer reviews," in 10th ACM SIGKDD international conference on Knowledge discovery and data mining, Seattle, WA, USA, 2004.

²⁰ N. Jindal and B. Liu, "Opinion spam and analysis," Proceedings of the 2008 International Conference on Web Search and Data Mining, ACM, pp. 219-230, 2008.

²¹ A. Tumasjan, T. O. Sprenger, P. G. Sandner and I. M. Welpe, "Predicting Elections with Twitter: What 140 Characters Reveal about Political Sentiment.," ICWSM, vol. 10, pp. 178-185, 2010.

²² B. A. Huberman and S. Asur, "Predicting the Future with Social Media," in IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT), Toronto, ON, Canada, 2010.

²³ J. Bollen, H. Mao and X. Zeng, "Twitter mood predicts the stock market," Journal of Computational Science, vol. 2, no. 1, pp. 1-8, 2011.

²⁴ A. C. Madrigal, "The Atlantic," 18 March 2011. [Online]. Available: http://www.theatlantic.com/technology/archive/2011/03/does-anne-hathaway-news-drive-berkshire-hathaways-stock/72661/. [Accessed 06 June 2014].

²⁵ A. Celikyilmaz, D. Hakkani-Tur and J. Feng, "Probabilistic model-based sentiment analysis of twitter messages," Spoken Language Technology Workshop (SLT), IEEE, pp. 79 - 84, 2010.

²⁶ B. Pang and L. Lee, "Opinion Mining and Sentiment Analysis," Foundations and Trends in Information Retrieval, vol. 2, no. 1-2, pp. 1-135, 2008.

²⁷ Y. Choi and C. Cardie, "Learning with Compositional Semantics as Structural Inference for Subsentential Sentiment Analysis," in Proceedings of the Conference on Empirical Methods in Natural Language Processing, Honolulu, Hawaii, 2008.

²⁸ I. G. Councill, R. McDonald and L. Velikovich, "What's great and what's not: learning to classify the scope of negation for improved sentiment analysis," in Proceedings of the Workshop on Negation and Speculation in Natural Language Processing, Uppsala, Sweden, 2010.

²⁹ L. Jia, C. Yu and W. Meng, "The Effect of Negation on Sentiment Analysis and Retrieval Effectiveness," in Proceedings of the 18th ACM conference on Information and knowledge management, Hong Kong, China, 2009.

³⁰ P. Carvalho, L. Sarmento, M. J. Silva and E. de Oliveira, "Clues for detecting irony in user-generated contents: oh...!! it's "so easy" ;-)," in Proceedings of the 1st international CIKM workshop on Topic-sentiment analysis for mass opinion, Hong Kong, China, 2009.

³¹ M. Wiegand, A. Balahur, B. Roth, D. Klakow and A. Montoyo, "A survey on the role of negation in sentiment analysis," in Proceedings of the Workshop on Negation and Speculation in Natural Language Processing, Uppsala, Sweden, 2010.

³² J. Zhang and S. Tan, "An empirical study of sentiment analysis for chinese documents," Expert Systems with Applications, vol. 34, no. 4, pp. 2622-2629, 2008.

³³ K. Denecke, "Using SentiWordNet for multilingual sentiment analysis," in IEEE 24th International Conference on Data Engineering Workshop, Cancun, 2008.

³⁴ K. Hiroshi, N. Tetsuya and W. Hideo, "Deeper sentiment analysis using machine translation technology," in Proceedings of the 20th international conference on Computational Linguistics, 2004.

³⁵ M. Bautin, L. Vijayarenu and S. Skie, "International Sentiment Analysis for News and Blogs," in ICWSM, 2008.

³⁶ J. Boyd-Graber and P. Resnik , "Holistic sentiment analysis across languages: multilingual supervised latent Dirichlet allocation," in Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, Cambridge, Massachusetts, 2010.

³⁷ A. Abbasi, H. Chen and A. Salem , "Sentiment analysis in multiple languages: Feature selection for opinion classification in Web forums," ACM Trans. Inf. Syst., vol. 26, no. 3, pp. 1-34, 2008.

³⁸ J. B. Walther and K. P. D’Addario, "The impacts of emoticons on message interpretation in computer-mediated communication," Social science computer review, vol. 19, no. 3, pp. 324-347, 2001.

³⁹ J. Read , "Using emoticons to reduce dependency in machine learning techniques for sentiment classification," in Proceedings of the ACL Student Research Workshop, Ann Arbor, Michigan, 2005.

⁴⁰ A. Aue and M. Gamon, "Customizing Sentiment Classifiers to New Domains: a Case Study," in Proceedings of recent advances in natural language processing (RANLP), 2005.

⁴¹ A. Fahrni and M. Klenner, "Old wine or warm beer: Target-specific sentiment analysis of adjectives," in Proc. of the Symposium on Affective Language in Human and Machine, AISB, 2008.

⁴² F. Benamara, C. Cesarano, A. Picariello, D. Reforgiato and V. Subrahmanian, "Sentiment Analysis: Adjectives and Adverbs are better than Adjectives Alone," in ICWSM, 2007.

⁴³ M. S. Usha and M. Indra Devi, "Analysis of Sentiments using Unsupervised Learning Techniques," in International Conference on Information Communication and Embedded Systems (ICICES), Chennai, India, 2013.

⁴⁴ S. Tan, G. Wu, H. Tang and X. Cheng, "A Novel Scheme for Domain-transfer Problem in the context of Sentiment Analysis," in Proceedings of the sixteenth ACM conference on Conference on information and knowledge management , Lisbon, Portugal, 2007.