- Email: [email protected]
Challenges of Sentiment Analysis in Social Networks
CHAPTER
CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS: AN OVERVIEW
1
F.A. Pozzia , E. Fersinib , E. Messinab , B. Liuc SAS Institute Srl, Milan, Italya University of Milano-Bicocca, Milan, Italyb University of Illinois at Chicago, Chicago, IL, United Statesc
1 BACKGROUND Sentiment analysis, which is also called opinion mining, has been one of the most active research areas in natural language processing since early 2000 [1]. The aim of sentiment analysis is to define automatic tools able to extract subjective information from texts in natural language, such as opinions and sentiments, so as to create structured and actionable knowledge to be used by either a decision support system or a decision maker. Unsurprisingly, there has been some confusion among researchers about the difference between sentiment and opinion, thus debating whether the field should be called sentiment analysis or opinion mining. In Merriam-Webster’s Collegiate Dictionary, sentiment is defined as an attitude, thought, or judgment prompted by feeling, whereas opinion is defined as a view, judgment, or appraisal formed in the mind about a particular matter. The difference is quite subtle, and each of them contains some elements of the other. The definitions indicate that an opinion is more of a person’s concrete view about something, whereas a sentiment is more of a feeling. For example, the sentence “I am concerned about the current political situation” expresses a sentiment, whereas the sentence “I think politics is not doing well” expresses an opinion. If someone says the first sentence in a conversation, we can respond by saying “I share your sentiment,” but for the second sentence we would normally say “I agree/disagree with you.” However, the underlying meanings of the two sentences are strictly related because the sentiment depicted in the first sentence is likely to be a feeling caused by the opinion in the second sentence. Conversely, the first sentiment sentence implies a negative opinion about politics, which is what the second sentence is saying. Although in most cases opinions imply positive or negative sentiments, some opinions do not, such as “I think he will win at the next presidential election.” More formally, as defined in [1], an opinion is a quintuple, (ei , aij , sijkl , hk , tl ),
(1.1)
where ei is the name of an entity, aij is an aspect of ei , sijkl is the sentiment on aspect aij of entity ei , hk denotes the opinion holder, and tl is the time when the opinion is expressed by hk . Sentiment Analysis in Social Networks. http://dx.doi.org/10.1016/B978-0-12-804412-4.00001-2 Copyright © 2017 Elsevier Inc. All rights reserved.
1
2
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
2009
2010
2011
2012
2013
2014
2015
2016
FIG. 1.1 Google Trends data related to the keywords sentiment analysis.
The sentiment sijkl is positive, negative, or neutral, or expressed with different strength/intensity levels, such as the 1–5 stars system used by most review websites (eg, Amazon1 ). For example, consider that yesterday John bought an iPhone. He tested it during the whole day and when he went home from work (at 19:00 on 2-15-2014) he wrote on his favorite social network the message “The iPhone is very good, but they still need to work on battery life and security issues.” Let us index “iPhone,” “battery life,” and “security” as 1, 2, and 3 respectively. John is indexed as 4 and the time when he wrote the sentence is indexed as 5. Then John is the opinion holder h4 and t5 (“19:00 2-15-2014”) is the time when the opinion is expressed by h4 (John) . The term “iPhone” is the entity e1 , “battery life” and “security issues” are aspects a12 and a13 of entity e1 (“iPhone”), s1245 = neg is the sentiment on aspect a12 (“battery life”) of entity e1 (“iPhone”). and s1345 = neg is the sentiment on aspect a13 (“security issues”) of entity e1 (“iPhone’). When an opinion is on the entity itself as a whole, the special aspect “GENERAL” is used to denote it. From the definition of sentiment analysis reported above, “the aim of sentiment analysis is therefore to define automatic tools able to extract subjective information in order to create structured and actionable knowledge.” In line with this, the quintuple-based definition provides a framework to transform unstructured text to structured data (eg, a database table). Then a rich set of qualitative, quantitative, and trend analyses can be performed with traditional database management systems and online analytical processing tools. Because of the importance of sentiment analysis to business and society, it has spread from computer science to management science and the social sciences. In recent years industrial activities surrounding sentiment analysis have also thrived: numerous start-ups have emerged, and many large corporations have built their own in-house capabilities (eg, Microsoft, Google, Hewlett-Packard, IBM, SAP, and SAS Global Communications). Thanks to its strong applicability and interest in both the academic field and the industrial field, sentiment analysis is nowadays a trending topic. Fig. 1.1 represents the Google Trends data related to the keywords sentiment analysis, clearly demonstrating the continuous and increasing interest in this field. Nowadays, sentiment analysis has gained even more value with the advent of social networks. Their great diffusion and their role in modern society represent one of the most interesting novelties 1 http://www.amazon.com
1 BACKGROUND
3
in recent years, capturing the interest of researchers, journalists, companies, and governments. The dense interconnection that often arises among active users generates a discussion space that is able to motivate and involve individuals of a larger agora, linking people with common objectives and facilitating diverse forms of collective action. Social networks are therefore creating a digital revolution, enabling the expression and spread of emotions and opinions through the network, opening a window on others’ respective worlds, and snooping into their lives. Opinionated data on the net, if properly collected and analyzed, allow one not only to understand and explain many complex social phenomena but also to predict them. Considering that nowadays the current technological progress enables the efficient storing and retrieval of a huge amount of data, the current focus is now on methods for extracting information and creating knowledge from raw sources. Social networks represent an emerging challenging sector in the context of big data: the natural language expressions of people can be easily reported through short text messages, rapidly creating unique content of huge dimensions that must be efficiently and effectively analyzed to create actionable knowledge for decision making processes. The massive quantity of continuously contributing texts in social networks, which should be processed in real time so as to make informed decisions, calls for two main types of radical progress: (1) a change of direction in the research through the transition from a data-constrained to dataenabled paradigm and (2) the convergence to a multidisciplinary area that mainly takes advantage of psychology, sociology, natural language processing, and machine learning. The knowledge embedded in social network content has been shown to be of paramount importance from both user and company/organization points of view: while people express opinions on any kind of topic in an unconstrained and unbiased environment, corporations and institutions can gauge valuable information from raw sources. To make qualitative textual data effectively functional for decision processes, the quantification of “what people think” becomes a mandatory step. However, sentiment analysis is often improperly used when one is referring to polarity classification, which instead is a subtask aimed at extracting positive, negative, or neutral sentiments (also called polarities) from texts. Although an opinion could also have a neutral polarity (eg, “I don’t know if I liked the movie or not. I should watch it quietly.”), most work in sentiment analysis usually assumes only positive and negative sentiments for simplicity. Depending on the field of application, several names are used for sentiment analysis (eg, opinion mining, opinion extraction, sentiment mining, subjectivity analysis, affect analysis, emotion analysis, and review mining). A taxonomy of the most popular sentiment analysis tasks is reported in Fig. 1.2. Sentiment Analysis in Social Networks tries to overcome this limitation by (1) collecting and proposing new relevant research work from experts in the field, (2) debating the advantages and disadvantages when one is applying sentiment analysis in social networks, and (3) discussing the progress of sentiment analysis in social networks and future directions. This book will accurately investigate the above-mentioned needs by providing advanced and specific solutions to address sentiment analysis in social networks. In particular, it presents the latest work by some of the most relevant experts in the field. At the end, a detailed conclusive discussion is provided and the personal and valuable thoughts and opinions of these researchers on future directions are presented. Although polarity classification is usually considered a core task because of its direct utility and applicability in working systems, all the chapters aspire to be relevant with respect to the needs outlined above. This is not a mere protocol declaration: the book has been thought about and designed
4
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
Sentiment analysis
Aspect-based sentiment classification
Subjectivity classification
Contextual polarity disambiguation
Polarity classification
Sentiment rating prediction
Opinion summarization
Cross-domain sentiment classification
Opinion visualization
Cross-language sentiment classification
Sarcasm detection
Entity, opinion holder and time extraction Coreference resolution and word sense disambiguation Sentiment lexicon generation Opinion search and retrieval
Opinion spam detection
FIG. 1.2 Sentiment analysis tasks.
as a whole, as an indivisible gold mine, which intends to provide contributions highly connected to each other.
2 SENTIMENT ANALYSIS IN SOCIAL NETWORKS: A NEW RESEARCH APPROACH The general trend in research regarding sentiment analysis in social networks is to apply the techniques inherited from traditional sentiment analysis studied since early 2000. However, considering the
3 SENTIMENT ANALYSIS CHARACTERISTICS
5
evolution of the sources where opinions are voiced, the strategies available in the current state of the art are no longer effective for mining opinions in this new and challenging environment. In fact, social network sentiment analysis, in addition to inheriting a multitude of issues from traditional sentiment analysis and natural language processing, introduces further complexities (short messages, noisy content, metadata such as gender, location, and age) and new sources of information not leveraged in traditional approaches. In particular, given that social networks are clearly having an impact on language, the daily challenges regarding sentiment analysis mainly focus on the constant evolution of the language used online in user-generated content: the words that surround us every day influence the words we use. Since much of the written language we see is now on the screens of our computers, tablets, and smartphones, language now evolves partly through our interaction with technology. And because the language used in social networks for us to communicate with each other tends to be more malleable than formal writing, the combination of informal, personal communication, and the mass audience afforded by social networks is a recipe for rapid change. Taking into serious consideration the continuous language revolution, we believe sentiment analysis systems should be able to natively adapt to it, or alternatively be adapted by researchers. Being able to juggle these problems requires strong natural language processing and linguistics skills. As a side effect, this language evolution strongly influences the way in which irony and sarcasm is uttered. A further daily challenge relates to the nature of social networks, which by definition are dynamic and heterogeneous and the entities involved are connected to each other. Conversely, a representation of real-world data where instances are considered as homogeneous, independent, and identically distributed leads us to a substantial loss of information and to the introduction of a statistical bias. Dealing with relational environments by our taking advantage of social network analysis becomes a mandatory step to go beyond the current state of the art, where only textual content is tackled. For this reason, the combination of content and relationships is a core task of the recent literature on sentiment analysis. A final crucial issue, which is usually overlooked, is concerned with visualization and summarization of opinions. This issue becomes more important when opinions need to be concisely presented over large networked environments. Traditional visual analytic tools need to be redesigned according to this novel necessity.
3 SENTIMENT ANALYSIS CHARACTERISTICS Sentiment analysis is a broad and complex field of research. In the following, the main characteristics that constitute sentiment analysis are described and discussed in detail.
3.1 SENTIMENT CATEGORIZATION: OBJECTIVE VERSUS SUBJECTIVE SENTENCES The first aim when one is dealing with sentiment analysis usually consists in distinguishing between subjective and objective sentences. If a given sentence is classified as objective, no other fundamental tasks are required, while if the sentence is classified as subjective, its polarity (positive, negative, or neutral) needs to be estimated (see Fig. 1.3). Subjectivity classification [2] is the task that distinguishes
6
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
Sentence Subjectivity classification : Objective Polarity classification :
Subjective
Positive Negative Neutral
FIG. 1.3 Sentiment analysis workflow.
sentences that express objective (or factual) information (objective sentences) from sentences that express subjective views and opinions (subjective sentences). An example of an objective sentence is “The iPhone is a smartphone,” while an example of a subjective sentence is “The iPhone is awesome.” Polarity classification is the task that distinguishes sentences that express positive, negative, or neutral polarities. Note that a subjective sentence may not express any positive or negative sentiment (eg, “I guess he has arrived”). For this reason, it should be classified as “neutral.”
3.2 LEVELS OF ANALYSIS As mentioned earlier, the aim of sentiment analysis is to “define automatic tools able to extract subjective information from texts in natural language.” The first choice when one is applying sentiment analysis is to define what text (ie, the analyzed object) means in the case of study considered. In general, sentiment analysis in social networks can be investigated mainly at three levels (represented graphically in Fig. 1.4):
FIG. 1.4 Different levels of analysis.
3 SENTIMENT ANALYSIS CHARACTERISTICS
•
•
•
7
Message level: The aim is to classify the polarity of a whole opinionated message. For example, given a product review, the system determines whether the text message expresses an overall positive, negative, or neutral opinion about the product. The assumption is that the entire message expresses only one opinion on a single entity (eg, a single product). Sentence level: The aim is to determine the polarity of each sentence contained in a text message. The assumption is that each sentence, in a given message, denotes a single opinion on a single entity. Entity and aspect level: Performs a finer-grained analysis than message and sentence level. It is based on the idea that an opinion consists of a sentiment and a target (of opinion). For example, the sentence “The iPhone is very good, but they still need to work on battery life and security issues” evaluates three aspects: iPhone (positive), battery life (negative), and security (negative).
3.3 REGULAR VERSUS COMPARATIVE OPINION An opinion can assume different shades and can be assigned to one of the following groups: •
•
Regular opinion: A regular opinion is often referred to in the literature as a standard opinion and it has two main subtypes: – Direct opinion: A direct opinion refers to an opinion expressed directly on an entity (eg, “The screen brightness of the iPhone is awesome”). – Indirect opinion: An indirect opinion is an opinion that is expressed indirectly on an entity on the basis of its effects on some other entities. For example, the sentence “After I switched to the iPhone, I lost all my data!” describes an undesirable effect of the switch on “my data,” which indirectly gives a negative sentiment to the iPhone. Comparative opinion: A comparative opinion expresses a relation of similarities or differences between two or more entities and/or a preference of the opinion holder based on some shared aspects of the entities [3]. For example, the sentences, “iOS is better performing than Android” and “iOS is the best performing operating system” express two comparative opinions. A comparative opinion is usually expressed with use of the comparative or superlative form of an adjective or adverb.
3.4 EXPLICIT VERSUS IMPLICIT OPINIONS Among the different shades that an opinion can assume, we have to distinguish explicit and implicit opinions: • •
Explicit opinion: An explicit opinion is a subjective statement that gives a regular or comparative opinion (eg, “The screen brightness of the iPhone is awesome”). Implicit opinion: An implicit opinion is an objective statement that implies a regular or comparative opinion that usually expresses a desirable or undesirable fact (eg, “Saturday night I’ll go to the movie theater to watch ‘Lone Survivor.’ I cannot wait to watch it!” and “‘Saving Private Ryan’ is more violent than ‘Lone Survivor’”). The first example suggests that there is some good expectation about the movie, although it is not explicated in words, while understanding the hidden opinion in the second example is difficult even for humans. For some people, violence in war
8
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
movies could be a good characteristic that makes the movie more realistic, whereas it could be a negative feature for others. Clearly, explicit opinions are easier to detect and to classify than implicit opinions. Much of the current research has focused on explicit opinions. Relatively less work has been done on implicit opinions.
3.5 THE ROLE OF SEMANTICS The semantics of the language used in social networks is fundamental to accurately analyze user expressions. The context of a textual expression is therefore a crucial element that should be taken into account to properly deal with the underlying sentiment. A sentence “taken as it is” can appear as negative or positive, but if it is properly analyzed from a semantic point of view it can be completely different. For instance, the sentences “I watched the most terrific horror movie. It was like a real nightmare! PAAAANIIIICCC” can be initially interpreted as negative, but taking into account the context where these kinds of opinions are expressed (ie, a community of horror-movie lovers) and some lexical cues that are typical of the social network language, we should derive a (real) positive judgment. Lexica, corpora, and ontologies need to be properly constructed and used for us to have a deep understanding of the semantics of the natural language in online social networks.
3.6 DEALING WITH FIGURES OF SPEECH A figure of speech is any artful deviation from the ordinary mode of speaking or writing [4]. In the tradition of Aristotle, figures of speech can be divided into two groups: schemes and tropes. The function of schemes and tropes is to carry out a transference of some kind; schemes are characterized by a transference in order, while tropes are characterized by a transference in meaning. For example, the most problematic figures of speech in natural language processing are irony and sarcasm, which are collocated under the tropes group. While irony is often used to emphasize occurrences that deviate from the expected, such as twists of fate, sarcasm is commonly used to convey implicit criticism with a particular victim as its target [5]. Examples of sarcastic and ironic sentences are: 1. Sarcasm (Note: Alice hates Bill’s travel books) – Alice: Yeah, I like, really dig your travel books, Bill. You’re a really skillful author. – Bill: Oh. 2. Irony (Note: Bill and Alice have just seen a really appalling play at the theater. Both Bill and Alice are disappointed.) – Bill: Well! What a worthwhile use of an evening! – Alice : Yeah. In the irony example, there was no sarcasm because Bill was not intending to wound Alice with his comment. He was using irony to remark that he felt he had wasted his evening at the theater. In the sarcasm example, Alice used sarcasm to show Bill that she did not like his books and thought that he is not a good writer. There is irony too, but the tone of the delivery that conveys implicit criticism makes it sarcastic.
4 APPLICATIONS
9
One inherent characteristic of the sarcastic and irony speech acts is that they are sometimes hard to recognize, first for humans and then for machines. The difficulty in the recognition of sarcasm and irony causes misunderstanding in everyday communication and poses problems to many natural language processing systems because of the poor results obtained by state-of-the-art work. In the context of sentiment analysis (where sarcasm and irony are usually considered as synonyms) when a sarcastic/ironic sentence is detected as positive, it likely means negative, and vice versa.
3.7 RELATIONSHIPS IN SOCIAL NETWORKS Sentiment analysis in social networks is generally based on the assumption that the texts provided by the users are independent and identically distributed. Although much effort has been expended on handling the complex characteristics of the language in social networking environments, consideration of user-generated content as networked text is still an open issue. A first tentative approach to deal with the real nature of social network content is related to the principle of homophily [6]. In this context, “friendship” relationships can be used to infer that connected users may be likelier to hold similar opinions. However, a sentiment analysis system should take into account that the assumption about the friendship relations does not properly reflect the real world, where two connected users could have different opinions about the same topic. According to this remark, several other pieces of relational information can be extracted from the social network itself for better representation of user and post connections. Relationships based on sharing activities or that represent an appreciation can be more informative than a simple friendship.
4 APPLICATIONS One of the most important needs of businesses and organizations in the real world is to find and analyze consumer or public opinions about their products and services (eg, “Why aren’t consumers buying our laptop?”). Knowing the opinions of existing users regarding a specific product is also interesting for individual consumers. This information could be useful to decide whether to buy the product or not. This shows that decision making processes are also common in everyday lives. However, with the advent of sentiment analysis, an individual is no longer strictly limited to asking friends and family for their opinions or an organization is no longer limited to conducting surveys, opinion polls, and focus groups to sound out public or consumer opinions. Sentiment analysis paves the way to several and interesting applications, in almost every possible domain. For example, summarizing user reviews is a relevant task. In addition, errors in user ratings could be fixed [7]: it is possible that users accidentally select a low rating when their review indicates a positive evaluation. Moreover, opinions matter a great deal in politics. Some work has focused on understanding what voters are thinking [8,9]. For instance, the US president Barack Obama used sentiment analysis to gauge the feelings of core voters during the 2008 presidential election. Other projects have as a long-term goal the clarification of politicians’ positions, such as what public figures support or oppose, to enhance the quality of information that voters have access to [10,11]. A further task is the augmentation of recommendation systems, where the system might not recommend items that receive negative feedback several times.
10
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
Moreover, ads are displayed in sidebars in some online systems. It could be useful to detect webpages that contain inappropriate content for the placement of ads [12]. The system could highlight product ads when relevant positive sentiments are detected, and hide the ads when negative statements are discovered. Opinionated documents could also have the form of organizations’ internal data (eg, customer feedback). Sentiment analysis applications have spread to several domains, from services and health care to financial services and political elections. However, sentiment analysis can also be applied to more ethical principles. For example, on the basis of observations of Twitter’s role in civilian response during the 2009 Jakarta and Mumbai terrorist attacks, Cheong and Lee [13] proposed a structured framework to harvest civilian sentiment and response on Twitter during terrorism scenarios. Coupled with intelligent data mining, visualization, and filtering methods, these data can be collated into a knowledge base that would be of great utility to decision makers and the authorities for rapid response and monitoring during such scenarios. Sentiment analysis is also applied to the medical field. Cobb et al. [14] applied sentiment analysis to examine how exposure to messages about the smokingcessation drug varenicline (used to treat nicotine addiction) affects smokers’ decision making regarding its use.
REFERENCES [1] B. Liu, Sentiment Analysis and Opinion Mining, Morgan & Claypool, San Rafael, CA, 2012. [2] J.M. Wiebe, R.F. Bruce, T.P. O’Hara, Development and use of a gold-standard data set for subjectivity classifications, in: Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics on Computational Linguistics, ACL ’99, Association for Computational Linguistics, 1999, pp. 246–253. [3] N. Jindal, B. Liu, Identifying comparative sentences in text documents, in: Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’06, ACM, 2006, pp. 244–251. [4] E.P.J. Corbett, Classical Rhetoric for the Modern Student, second ed., Oxford University Press, Oxford, UK, 1971. [5] S. McDonald, Exploring the process of inference generation in sarcasm: a review of normal and clinical studies, Brain Lang. 68 (3) (1999) 486–506. [6] P.F. Lazarsfeld, R.K. Merton, Friendship as a social process: a substantive and methodological analysis, in: M. Berger, T. Abel, C.H. Page (Eds.), Freedom and Control in Modern Society, Van Nostrand, New York, 1954, pp. 8–66. [7] B. Pang, L. Lee, Opinion Mining and Sentiment Analysis, Found. Trends Inf. Retr. 2 (1–2) (2008) 1–135. [8] A.B. Goldberg, X. Zhu, S.J. Wright, Dissimilarity in graph-based semi-supervised classification, in: AISTATS, vol. 2, 2007, pp. 155–162. [9] D. Hopkins, G. King, Extracting systematic social science meaning from text, Manuscript available at http:// gking.harvard.edu/files/words.pdf, (2007). [10] M. Bansal, C. Cardie, L. Lee, The power of negative thinking: exploiting label disagreement in the min-cut classification framework, in: COLING (Posters), 2008, pp. 15–18. [11] S.C. Greene, Spin: Lexical Semantics, Transitivity, and the Identification of Implicit Sentiment, ProQuest, Ann Arbor, MI, 2007.
REFERENCES
11
[12] X. Jin,Y. Li, T. Mah, J. Tong, Sensitive webpage classification for content advertising, in: Proceedings of the First International Workshop on Data Mining and Audience Intelligence for Advertising, ADKDD ’07, ACM, 2007, pp. 28–33. [13] M. Cheong, V.C.S. Lee, A microblogging-based approach to terrorism informatics: exploration and chronicling civilian sentiment and response to terrorism events via Twitter, Inform. Syst. Front. 13 (1) (2011) 45–59. [14] N.K. Cobb, D. Mays, A.L. Graham, Sentiment analysis to determine the impact of online messages on smokers’ choices to use varenicline, J. Natl. Cancer Inst. Monogr. 47 (2013) 224–230.
CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS: AN OVERVIEW
1
F.A. Pozzia , E. Fersinib , E. Messinab , B. Liuc SAS Institute Srl, Milan, Italya University of Milano-Bicocca, Milan, Italyb University of Illinois at Chicago, Chicago, IL, United Statesc
1 BACKGROUND Sentiment analysis, which is also called opinion mining, has been one of the most active research areas in natural language processing since early 2000 [1]. The aim of sentiment analysis is to define automatic tools able to extract subjective information from texts in natural language, such as opinions and sentiments, so as to create structured and actionable knowledge to be used by either a decision support system or a decision maker. Unsurprisingly, there has been some confusion among researchers about the difference between sentiment and opinion, thus debating whether the field should be called sentiment analysis or opinion mining. In Merriam-Webster’s Collegiate Dictionary, sentiment is defined as an attitude, thought, or judgment prompted by feeling, whereas opinion is defined as a view, judgment, or appraisal formed in the mind about a particular matter. The difference is quite subtle, and each of them contains some elements of the other. The definitions indicate that an opinion is more of a person’s concrete view about something, whereas a sentiment is more of a feeling. For example, the sentence “I am concerned about the current political situation” expresses a sentiment, whereas the sentence “I think politics is not doing well” expresses an opinion. If someone says the first sentence in a conversation, we can respond by saying “I share your sentiment,” but for the second sentence we would normally say “I agree/disagree with you.” However, the underlying meanings of the two sentences are strictly related because the sentiment depicted in the first sentence is likely to be a feeling caused by the opinion in the second sentence. Conversely, the first sentiment sentence implies a negative opinion about politics, which is what the second sentence is saying. Although in most cases opinions imply positive or negative sentiments, some opinions do not, such as “I think he will win at the next presidential election.” More formally, as defined in [1], an opinion is a quintuple, (ei , aij , sijkl , hk , tl ),
(1.1)
where ei is the name of an entity, aij is an aspect of ei , sijkl is the sentiment on aspect aij of entity ei , hk denotes the opinion holder, and tl is the time when the opinion is expressed by hk . Sentiment Analysis in Social Networks. http://dx.doi.org/10.1016/B978-0-12-804412-4.00001-2 Copyright © 2017 Elsevier Inc. All rights reserved.
1
2
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
2009
2010
2011
2012
2013
2014
2015
2016
FIG. 1.1 Google Trends data related to the keywords sentiment analysis.
The sentiment sijkl is positive, negative, or neutral, or expressed with different strength/intensity levels, such as the 1–5 stars system used by most review websites (eg, Amazon1 ). For example, consider that yesterday John bought an iPhone. He tested it during the whole day and when he went home from work (at 19:00 on 2-15-2014) he wrote on his favorite social network the message “The iPhone is very good, but they still need to work on battery life and security issues.” Let us index “iPhone,” “battery life,” and “security” as 1, 2, and 3 respectively. John is indexed as 4 and the time when he wrote the sentence is indexed as 5. Then John is the opinion holder h4 and t5 (“19:00 2-15-2014”) is the time when the opinion is expressed by h4 (John) . The term “iPhone” is the entity e1 , “battery life” and “security issues” are aspects a12 and a13 of entity e1 (“iPhone”), s1245 = neg is the sentiment on aspect a12 (“battery life”) of entity e1 (“iPhone”). and s1345 = neg is the sentiment on aspect a13 (“security issues”) of entity e1 (“iPhone’). When an opinion is on the entity itself as a whole, the special aspect “GENERAL” is used to denote it. From the definition of sentiment analysis reported above, “the aim of sentiment analysis is therefore to define automatic tools able to extract subjective information in order to create structured and actionable knowledge.” In line with this, the quintuple-based definition provides a framework to transform unstructured text to structured data (eg, a database table). Then a rich set of qualitative, quantitative, and trend analyses can be performed with traditional database management systems and online analytical processing tools. Because of the importance of sentiment analysis to business and society, it has spread from computer science to management science and the social sciences. In recent years industrial activities surrounding sentiment analysis have also thrived: numerous start-ups have emerged, and many large corporations have built their own in-house capabilities (eg, Microsoft, Google, Hewlett-Packard, IBM, SAP, and SAS Global Communications). Thanks to its strong applicability and interest in both the academic field and the industrial field, sentiment analysis is nowadays a trending topic. Fig. 1.1 represents the Google Trends data related to the keywords sentiment analysis, clearly demonstrating the continuous and increasing interest in this field. Nowadays, sentiment analysis has gained even more value with the advent of social networks. Their great diffusion and their role in modern society represent one of the most interesting novelties 1 http://www.amazon.com
1 BACKGROUND
3
in recent years, capturing the interest of researchers, journalists, companies, and governments. The dense interconnection that often arises among active users generates a discussion space that is able to motivate and involve individuals of a larger agora, linking people with common objectives and facilitating diverse forms of collective action. Social networks are therefore creating a digital revolution, enabling the expression and spread of emotions and opinions through the network, opening a window on others’ respective worlds, and snooping into their lives. Opinionated data on the net, if properly collected and analyzed, allow one not only to understand and explain many complex social phenomena but also to predict them. Considering that nowadays the current technological progress enables the efficient storing and retrieval of a huge amount of data, the current focus is now on methods for extracting information and creating knowledge from raw sources. Social networks represent an emerging challenging sector in the context of big data: the natural language expressions of people can be easily reported through short text messages, rapidly creating unique content of huge dimensions that must be efficiently and effectively analyzed to create actionable knowledge for decision making processes. The massive quantity of continuously contributing texts in social networks, which should be processed in real time so as to make informed decisions, calls for two main types of radical progress: (1) a change of direction in the research through the transition from a data-constrained to dataenabled paradigm and (2) the convergence to a multidisciplinary area that mainly takes advantage of psychology, sociology, natural language processing, and machine learning. The knowledge embedded in social network content has been shown to be of paramount importance from both user and company/organization points of view: while people express opinions on any kind of topic in an unconstrained and unbiased environment, corporations and institutions can gauge valuable information from raw sources. To make qualitative textual data effectively functional for decision processes, the quantification of “what people think” becomes a mandatory step. However, sentiment analysis is often improperly used when one is referring to polarity classification, which instead is a subtask aimed at extracting positive, negative, or neutral sentiments (also called polarities) from texts. Although an opinion could also have a neutral polarity (eg, “I don’t know if I liked the movie or not. I should watch it quietly.”), most work in sentiment analysis usually assumes only positive and negative sentiments for simplicity. Depending on the field of application, several names are used for sentiment analysis (eg, opinion mining, opinion extraction, sentiment mining, subjectivity analysis, affect analysis, emotion analysis, and review mining). A taxonomy of the most popular sentiment analysis tasks is reported in Fig. 1.2. Sentiment Analysis in Social Networks tries to overcome this limitation by (1) collecting and proposing new relevant research work from experts in the field, (2) debating the advantages and disadvantages when one is applying sentiment analysis in social networks, and (3) discussing the progress of sentiment analysis in social networks and future directions. This book will accurately investigate the above-mentioned needs by providing advanced and specific solutions to address sentiment analysis in social networks. In particular, it presents the latest work by some of the most relevant experts in the field. At the end, a detailed conclusive discussion is provided and the personal and valuable thoughts and opinions of these researchers on future directions are presented. Although polarity classification is usually considered a core task because of its direct utility and applicability in working systems, all the chapters aspire to be relevant with respect to the needs outlined above. This is not a mere protocol declaration: the book has been thought about and designed
4
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
Sentiment analysis
Aspect-based sentiment classification
Subjectivity classification
Contextual polarity disambiguation
Polarity classification
Sentiment rating prediction
Opinion summarization
Cross-domain sentiment classification
Opinion visualization
Cross-language sentiment classification
Sarcasm detection
Entity, opinion holder and time extraction Coreference resolution and word sense disambiguation Sentiment lexicon generation Opinion search and retrieval
Opinion spam detection
FIG. 1.2 Sentiment analysis tasks.
as a whole, as an indivisible gold mine, which intends to provide contributions highly connected to each other.
2 SENTIMENT ANALYSIS IN SOCIAL NETWORKS: A NEW RESEARCH APPROACH The general trend in research regarding sentiment analysis in social networks is to apply the techniques inherited from traditional sentiment analysis studied since early 2000. However, considering the
3 SENTIMENT ANALYSIS CHARACTERISTICS
5
evolution of the sources where opinions are voiced, the strategies available in the current state of the art are no longer effective for mining opinions in this new and challenging environment. In fact, social network sentiment analysis, in addition to inheriting a multitude of issues from traditional sentiment analysis and natural language processing, introduces further complexities (short messages, noisy content, metadata such as gender, location, and age) and new sources of information not leveraged in traditional approaches. In particular, given that social networks are clearly having an impact on language, the daily challenges regarding sentiment analysis mainly focus on the constant evolution of the language used online in user-generated content: the words that surround us every day influence the words we use. Since much of the written language we see is now on the screens of our computers, tablets, and smartphones, language now evolves partly through our interaction with technology. And because the language used in social networks for us to communicate with each other tends to be more malleable than formal writing, the combination of informal, personal communication, and the mass audience afforded by social networks is a recipe for rapid change. Taking into serious consideration the continuous language revolution, we believe sentiment analysis systems should be able to natively adapt to it, or alternatively be adapted by researchers. Being able to juggle these problems requires strong natural language processing and linguistics skills. As a side effect, this language evolution strongly influences the way in which irony and sarcasm is uttered. A further daily challenge relates to the nature of social networks, which by definition are dynamic and heterogeneous and the entities involved are connected to each other. Conversely, a representation of real-world data where instances are considered as homogeneous, independent, and identically distributed leads us to a substantial loss of information and to the introduction of a statistical bias. Dealing with relational environments by our taking advantage of social network analysis becomes a mandatory step to go beyond the current state of the art, where only textual content is tackled. For this reason, the combination of content and relationships is a core task of the recent literature on sentiment analysis. A final crucial issue, which is usually overlooked, is concerned with visualization and summarization of opinions. This issue becomes more important when opinions need to be concisely presented over large networked environments. Traditional visual analytic tools need to be redesigned according to this novel necessity.
3 SENTIMENT ANALYSIS CHARACTERISTICS Sentiment analysis is a broad and complex field of research. In the following, the main characteristics that constitute sentiment analysis are described and discussed in detail.
3.1 SENTIMENT CATEGORIZATION: OBJECTIVE VERSUS SUBJECTIVE SENTENCES The first aim when one is dealing with sentiment analysis usually consists in distinguishing between subjective and objective sentences. If a given sentence is classified as objective, no other fundamental tasks are required, while if the sentence is classified as subjective, its polarity (positive, negative, or neutral) needs to be estimated (see Fig. 1.3). Subjectivity classification [2] is the task that distinguishes
6
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
Sentence Subjectivity classification : Objective Polarity classification :
Subjective
Positive Negative Neutral
FIG. 1.3 Sentiment analysis workflow.
sentences that express objective (or factual) information (objective sentences) from sentences that express subjective views and opinions (subjective sentences). An example of an objective sentence is “The iPhone is a smartphone,” while an example of a subjective sentence is “The iPhone is awesome.” Polarity classification is the task that distinguishes sentences that express positive, negative, or neutral polarities. Note that a subjective sentence may not express any positive or negative sentiment (eg, “I guess he has arrived”). For this reason, it should be classified as “neutral.”
3.2 LEVELS OF ANALYSIS As mentioned earlier, the aim of sentiment analysis is to “define automatic tools able to extract subjective information from texts in natural language.” The first choice when one is applying sentiment analysis is to define what text (ie, the analyzed object) means in the case of study considered. In general, sentiment analysis in social networks can be investigated mainly at three levels (represented graphically in Fig. 1.4):
FIG. 1.4 Different levels of analysis.
3 SENTIMENT ANALYSIS CHARACTERISTICS
•
•
•
7
Message level: The aim is to classify the polarity of a whole opinionated message. For example, given a product review, the system determines whether the text message expresses an overall positive, negative, or neutral opinion about the product. The assumption is that the entire message expresses only one opinion on a single entity (eg, a single product). Sentence level: The aim is to determine the polarity of each sentence contained in a text message. The assumption is that each sentence, in a given message, denotes a single opinion on a single entity. Entity and aspect level: Performs a finer-grained analysis than message and sentence level. It is based on the idea that an opinion consists of a sentiment and a target (of opinion). For example, the sentence “The iPhone is very good, but they still need to work on battery life and security issues” evaluates three aspects: iPhone (positive), battery life (negative), and security (negative).
3.3 REGULAR VERSUS COMPARATIVE OPINION An opinion can assume different shades and can be assigned to one of the following groups: •
•
Regular opinion: A regular opinion is often referred to in the literature as a standard opinion and it has two main subtypes: – Direct opinion: A direct opinion refers to an opinion expressed directly on an entity (eg, “The screen brightness of the iPhone is awesome”). – Indirect opinion: An indirect opinion is an opinion that is expressed indirectly on an entity on the basis of its effects on some other entities. For example, the sentence “After I switched to the iPhone, I lost all my data!” describes an undesirable effect of the switch on “my data,” which indirectly gives a negative sentiment to the iPhone. Comparative opinion: A comparative opinion expresses a relation of similarities or differences between two or more entities and/or a preference of the opinion holder based on some shared aspects of the entities [3]. For example, the sentences, “iOS is better performing than Android” and “iOS is the best performing operating system” express two comparative opinions. A comparative opinion is usually expressed with use of the comparative or superlative form of an adjective or adverb.
3.4 EXPLICIT VERSUS IMPLICIT OPINIONS Among the different shades that an opinion can assume, we have to distinguish explicit and implicit opinions: • •
Explicit opinion: An explicit opinion is a subjective statement that gives a regular or comparative opinion (eg, “The screen brightness of the iPhone is awesome”). Implicit opinion: An implicit opinion is an objective statement that implies a regular or comparative opinion that usually expresses a desirable or undesirable fact (eg, “Saturday night I’ll go to the movie theater to watch ‘Lone Survivor.’ I cannot wait to watch it!” and “‘Saving Private Ryan’ is more violent than ‘Lone Survivor’”). The first example suggests that there is some good expectation about the movie, although it is not explicated in words, while understanding the hidden opinion in the second example is difficult even for humans. For some people, violence in war
8
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
movies could be a good characteristic that makes the movie more realistic, whereas it could be a negative feature for others. Clearly, explicit opinions are easier to detect and to classify than implicit opinions. Much of the current research has focused on explicit opinions. Relatively less work has been done on implicit opinions.
3.5 THE ROLE OF SEMANTICS The semantics of the language used in social networks is fundamental to accurately analyze user expressions. The context of a textual expression is therefore a crucial element that should be taken into account to properly deal with the underlying sentiment. A sentence “taken as it is” can appear as negative or positive, but if it is properly analyzed from a semantic point of view it can be completely different. For instance, the sentences “I watched the most terrific horror movie. It was like a real nightmare! PAAAANIIIICCC” can be initially interpreted as negative, but taking into account the context where these kinds of opinions are expressed (ie, a community of horror-movie lovers) and some lexical cues that are typical of the social network language, we should derive a (real) positive judgment. Lexica, corpora, and ontologies need to be properly constructed and used for us to have a deep understanding of the semantics of the natural language in online social networks.
3.6 DEALING WITH FIGURES OF SPEECH A figure of speech is any artful deviation from the ordinary mode of speaking or writing [4]. In the tradition of Aristotle, figures of speech can be divided into two groups: schemes and tropes. The function of schemes and tropes is to carry out a transference of some kind; schemes are characterized by a transference in order, while tropes are characterized by a transference in meaning. For example, the most problematic figures of speech in natural language processing are irony and sarcasm, which are collocated under the tropes group. While irony is often used to emphasize occurrences that deviate from the expected, such as twists of fate, sarcasm is commonly used to convey implicit criticism with a particular victim as its target [5]. Examples of sarcastic and ironic sentences are: 1. Sarcasm (Note: Alice hates Bill’s travel books) – Alice: Yeah, I like, really dig your travel books, Bill. You’re a really skillful author. – Bill: Oh. 2. Irony (Note: Bill and Alice have just seen a really appalling play at the theater. Both Bill and Alice are disappointed.) – Bill: Well! What a worthwhile use of an evening! – Alice : Yeah. In the irony example, there was no sarcasm because Bill was not intending to wound Alice with his comment. He was using irony to remark that he felt he had wasted his evening at the theater. In the sarcasm example, Alice used sarcasm to show Bill that she did not like his books and thought that he is not a good writer. There is irony too, but the tone of the delivery that conveys implicit criticism makes it sarcastic.
4 APPLICATIONS
9
One inherent characteristic of the sarcastic and irony speech acts is that they are sometimes hard to recognize, first for humans and then for machines. The difficulty in the recognition of sarcasm and irony causes misunderstanding in everyday communication and poses problems to many natural language processing systems because of the poor results obtained by state-of-the-art work. In the context of sentiment analysis (where sarcasm and irony are usually considered as synonyms) when a sarcastic/ironic sentence is detected as positive, it likely means negative, and vice versa.
3.7 RELATIONSHIPS IN SOCIAL NETWORKS Sentiment analysis in social networks is generally based on the assumption that the texts provided by the users are independent and identically distributed. Although much effort has been expended on handling the complex characteristics of the language in social networking environments, consideration of user-generated content as networked text is still an open issue. A first tentative approach to deal with the real nature of social network content is related to the principle of homophily [6]. In this context, “friendship” relationships can be used to infer that connected users may be likelier to hold similar opinions. However, a sentiment analysis system should take into account that the assumption about the friendship relations does not properly reflect the real world, where two connected users could have different opinions about the same topic. According to this remark, several other pieces of relational information can be extracted from the social network itself for better representation of user and post connections. Relationships based on sharing activities or that represent an appreciation can be more informative than a simple friendship.
4 APPLICATIONS One of the most important needs of businesses and organizations in the real world is to find and analyze consumer or public opinions about their products and services (eg, “Why aren’t consumers buying our laptop?”). Knowing the opinions of existing users regarding a specific product is also interesting for individual consumers. This information could be useful to decide whether to buy the product or not. This shows that decision making processes are also common in everyday lives. However, with the advent of sentiment analysis, an individual is no longer strictly limited to asking friends and family for their opinions or an organization is no longer limited to conducting surveys, opinion polls, and focus groups to sound out public or consumer opinions. Sentiment analysis paves the way to several and interesting applications, in almost every possible domain. For example, summarizing user reviews is a relevant task. In addition, errors in user ratings could be fixed [7]: it is possible that users accidentally select a low rating when their review indicates a positive evaluation. Moreover, opinions matter a great deal in politics. Some work has focused on understanding what voters are thinking [8,9]. For instance, the US president Barack Obama used sentiment analysis to gauge the feelings of core voters during the 2008 presidential election. Other projects have as a long-term goal the clarification of politicians’ positions, such as what public figures support or oppose, to enhance the quality of information that voters have access to [10,11]. A further task is the augmentation of recommendation systems, where the system might not recommend items that receive negative feedback several times.
10
CHAPTER 1 CHALLENGES OF SENTIMENT ANALYSIS IN SOCIAL NETWORKS
Moreover, ads are displayed in sidebars in some online systems. It could be useful to detect webpages that contain inappropriate content for the placement of ads [12]. The system could highlight product ads when relevant positive sentiments are detected, and hide the ads when negative statements are discovered. Opinionated documents could also have the form of organizations’ internal data (eg, customer feedback). Sentiment analysis applications have spread to several domains, from services and health care to financial services and political elections. However, sentiment analysis can also be applied to more ethical principles. For example, on the basis of observations of Twitter’s role in civilian response during the 2009 Jakarta and Mumbai terrorist attacks, Cheong and Lee [13] proposed a structured framework to harvest civilian sentiment and response on Twitter during terrorism scenarios. Coupled with intelligent data mining, visualization, and filtering methods, these data can be collated into a knowledge base that would be of great utility to decision makers and the authorities for rapid response and monitoring during such scenarios. Sentiment analysis is also applied to the medical field. Cobb et al. [14] applied sentiment analysis to examine how exposure to messages about the smokingcessation drug varenicline (used to treat nicotine addiction) affects smokers’ decision making regarding its use.
REFERENCES [1] B. Liu, Sentiment Analysis and Opinion Mining, Morgan & Claypool, San Rafael, CA, 2012. [2] J.M. Wiebe, R.F. Bruce, T.P. O’Hara, Development and use of a gold-standard data set for subjectivity classifications, in: Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics on Computational Linguistics, ACL ’99, Association for Computational Linguistics, 1999, pp. 246–253. [3] N. Jindal, B. Liu, Identifying comparative sentences in text documents, in: Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’06, ACM, 2006, pp. 244–251. [4] E.P.J. Corbett, Classical Rhetoric for the Modern Student, second ed., Oxford University Press, Oxford, UK, 1971. [5] S. McDonald, Exploring the process of inference generation in sarcasm: a review of normal and clinical studies, Brain Lang. 68 (3) (1999) 486–506. [6] P.F. Lazarsfeld, R.K. Merton, Friendship as a social process: a substantive and methodological analysis, in: M. Berger, T. Abel, C.H. Page (Eds.), Freedom and Control in Modern Society, Van Nostrand, New York, 1954, pp. 8–66. [7] B. Pang, L. Lee, Opinion Mining and Sentiment Analysis, Found. Trends Inf. Retr. 2 (1–2) (2008) 1–135. [8] A.B. Goldberg, X. Zhu, S.J. Wright, Dissimilarity in graph-based semi-supervised classification, in: AISTATS, vol. 2, 2007, pp. 155–162. [9] D. Hopkins, G. King, Extracting systematic social science meaning from text, Manuscript available at http:// gking.harvard.edu/files/words.pdf, (2007). [10] M. Bansal, C. Cardie, L. Lee, The power of negative thinking: exploiting label disagreement in the min-cut classification framework, in: COLING (Posters), 2008, pp. 15–18. [11] S.C. Greene, Spin: Lexical Semantics, Transitivity, and the Identification of Implicit Sentiment, ProQuest, Ann Arbor, MI, 2007.
REFERENCES
11
[12] X. Jin,Y. Li, T. Mah, J. Tong, Sensitive webpage classification for content advertising, in: Proceedings of the First International Workshop on Data Mining and Audience Intelligence for Advertising, ADKDD ’07, ACM, 2007, pp. 28–33. [13] M. Cheong, V.C.S. Lee, A microblogging-based approach to terrorism informatics: exploration and chronicling civilian sentiment and response to terrorism events via Twitter, Inform. Syst. Front. 13 (1) (2011) 45–59. [14] N.K. Cobb, D. Mays, A.L. Graham, Sentiment analysis to determine the impact of online messages on smokers’ choices to use varenicline, J. Natl. Cancer Inst. Monogr. 47 (2013) 224–230.