13

Big Social Data Analysis

Erik Cambria, Dheeraj Rajagopal, Daniel Olsher, and Dipankar Das CoNTeNTs

From Small to Big Social Data Analysis

As the Web rapidly evolves, Web users too are evolving with it. In an era of social connectedness, people are becoming more and more enthusiastic about interacting, sharing, and collaborating through social networks, online communities, blogs, Wikis, and other online collaborative media. In recent years, this collective intelligence has spread to many different areas, with particular focus on fields related to everyday life such as commerce, tourism, education, and health, causing the size of the social Web to expand exponentially. The distillation of knowledge from such a large amount of unstructured information, however, is an extremely difficult task, as the contents of todays Web are perfectly suitable for human consumption, but remain hardly accessible to machines. Big social data analysis grows out of this need and it includes disciplines such as social network analysis, multimedia management, social media analytics, trend discovery, and opinion mining. The opportunity to capture the opinions of the general public about social events, political movements, company strategies, marketing campaigns, and product preferences, in particular, has raised growing interest both within the scientific community, leading to many exciting open challenges, as well as in the business world, due to the remarkable benefits to be had from marketing and financial market prediction. This has led to the emerging fields of opinion mining and sentiment analysis, which deal with information retrieval and knowledge discovery from text using data mining and natural language processing (NLP) techniques to distill knowledge and opinions from the huge amount of information on

Pr oo

fs

fo rr

ev

ie w

From Small to Big Social Data Analysis............................................................ 401 Large-Scale Sentiment Analysis and Tracking................................................. 403 Toward the Concept-Level Analysis of Big Sentiment Data.......................... 406 Conclusion............................................................................................................ 412 References. ............................................................................................................. 412

K16832_C013.indd 401

on ly
401
9/5/2013 8:04:59 PM

402

Big Data Computing

the World Wide Web. Opinion mining and sentiment analysis are branches of the broad field of text data mining [21] and refer generally to the process of extracting interesting and nontrivial patterns or knowledge from unstructured text documents. They can be viewed as an extension of data mining or knowledge discovery from (structured) databases [12,36]. As the most natural form of storing information is text, opinion mining is believed to have a commercial potential higher than that of data mining. Opinion mining, however, is also a much more complex task, as it involves dealing with text data that are inherently big, unstructured, and fuzzy. Most of the existing approaches to opinion mining rely on the extraction of a vector representing the most salient and important text features, which is later used for classification purposes. Some of the most commonly used features are term frequency [42] and presence [33]. The latter is a binary-valued feature vectors in which the entries merely indicate whether a term occurs (value 1) or not (value 0) formed a more effective basis for reviewing polarity classification. This is indicative of an interesting difference between typical topic-based text categorization and polarity classification. While a topic is more likely to be emphasized by frequent occurrences of certain keywords, overall sentiment may not usually be highlighted through repeated use of the same terms. Other term-based features are often added to the features vector. Position is one of these, in consideration of how the position of a token in a text unit can affect the way in which the token affects the sentiment of the text. Also presence of n-grams, typically bi-grams and tri-grams, is often taken into account as useful features. Some methods also rely on the distance between terms. Parts of speech (POS) information (nouns, adjectives, adverbs, verbs, etc.) are also commonly exploited in general textual analysis as a basic form of word sense disambiguation [41]. Certain adjectives, in particular, have been proved to be good indicators of sentiment and sometimes have been used to guide feature selection for sentiment classification. In other works, eventually, the detection of sentiments was performed through selected phrases, which were chosen via a number of prespecified POS patterns, most including an adjective or an adverb [39]. All such approaches mainly rely on parts of text in which opinions and sentiments are explicitly expressed, for example, polarity terms, affect words, and their co-occurrence frequencies. Opinions and sentiments, however, are often conveyed implicitly through context- and domain-dependent concepts, which make purely syntactical approaches ineffective. To this end, novel approaches that go beyond mere word-level sentiment analysis are needed. Such approaches should employ new techniques capable of better grasping the conceptual rules that govern sentiment and the clues that can convey these concepts from realization to verbalization in the human mind. Next-generation opinion mining systems need broader and deeper common-sense knowledge bases and more cognitive and affective inspired reasoning methods, in order to better understand natural language opinions and sentiments and, hence, more

Pr oo

K16832_C013.indd 402

fs

fo rr

ev

ie w

on ly
9/5/2013 8:04:59 PM

Big Social Data Analysis

403

efficiently bridge the gap between (unstructured) textual information and (structured) machine-processable data.

Large-Scale Sentiment Analysis and Tracking

Sentiment analysis research at large scale has mainly been performed through syntactical approaches. Godbole etal. [16] proposed a system for large-scale analysis of news and blogs built on Lydia [32]. Their approach involves the identification of sentiments and the subsequent assignment of orientation to them. Seven sentiment dimensions were chosen, namely general, health, crime, sports, business, politics, and media. An initial seed of sentiments is expanded using the synonym and antonym queries in two steps. During the first iteration, a preliminary score for each word is calculated using a WordNet-based depth function. During the second iteration, the number of flips, or apparent alternations, of words is calculated. The words that qualify above the 5-hop threshold are classified as valid sentiment words. The extracted sentiments were assigned both a polarity and a subjectivity score, which were exploited for the classification of news and blogs. Kucuktunc etal. [24] showed that human factors affect sentiments extracted from Big Data. They analyzed the Yahoo! Answers Corpus, a large-scale online question and answering system, where people ask questions and the community provides relevant answers. Users can choose a category for the question and vote on answers. Corpus contained 412M sentences with about 34M questions and 132M answers. Several features that affected the sentiment of answers were analyzed: textual features, for example, question length (a general observation was that negative sentiments were dominant in the longer questions), punctuation (after 34 question marks, the attitude and sentimentality of the sentence shows a decreasing trend), and starting words; topical and temporal features; demographical features, for example, gender (women show a general positive attitude in their answers, whereas men are mostly neutral), age (a general trend is that the sentimentality decreases with increasing age), and education level (an increased sentimentality level is observed with increasing education level). A large-scale distributed system for Twitter sentiment analysis was built by Khuc etal. [23]. Twitter is a huge source of text data but the character limit of 140 makes it difficult to mine the sentiment data effectively. Twitter data, in fact, do not comply well with the existing NLP engines because of the increased use of nonlexical words and smileys to indicate sentiments. The primary advantage of Khuc etal.s system is the identification of sentiments of nontrivial sentiment words such as abbreviations and extended words, for example, lol and veeerrryyy goooodddd, and emoticons. After a lexicon builder

Pr oo

K16832_C013.indd 403

fs

fo rr

ev

ie w

on ly
9/5/2013 8:04:59 PM

404

Big Data Computing

normalizes the incoming tweets, a co-occurrence matrix is created through bi-gram phrases using the MapRreduce framework. The cosine similarity between words is then computed and the edges with low cosine score are removed. Finally, a sentiment score propagation algorithm is implemented and the sentiment scores are calculated. Brew etal. [2] use a crowdsourced approach to label sentiments. They ask volunteers to manually annotate RSS feeds and use these for training on a daily basis, in order to create a sentiment tag cloud. Although the approach presents very good accuracy, the scope of the work is very narrow. The major issue with using machine-learning techniques is that a model that works well for sentiments in one domain might not work well for another. To overcome such an issue, Glorot etal. [14] proposed a deep learning approach that trains a classifier in one domain and successfully applies it to another. This method is called domain adaptation. The intuition behind using the intermediate concepts is that these better guide the semantic and affective transfer among domains. The approach works at two levels. Firstly, an unsupervised classifier learns higher-level features from Amazon reviews in all available domains. Secondly, a linear SVM with squared hinge loss is used for the training phase on the labeled data from the source domain and tested on the target domain. A task that immediately derives from large-scale sentiment analysis is big sentiment data tracking, that is, the aggregation of all the scattered emotional information related to a specific topic through the delivery of an at-aglance presentation of the main sentiment topics or a graphical visualization of how sentiment changes from time to time over multiple events. To this end, related sentiment components need to be gathered and associated in a structured way, so that the affective information is tractable with respect to user, topic/event, sentiment, and time. In many applications, analysts and other users are interested in tracking changes in sentiment about a product, political candidate, company, or other issues, over time. Early work proposed to detect and track events from text archives [40]. Such events were extracted from different newspaper texts, stories, and other important documents, where occurrence time of events, temporal location, and ordering of the events were specified. The eventsentiment association highlights the fact that, though events and sentiments are closely coupled with each other from social, psychological, and commercial perspectives, there has been very little attention regarding their detection. The identification of the temporal relations between two events by taking the sentiment feature into account is also crucial in analyzing and tracking human sentiments. This is also important in a wide range of other NLP applications that include temporal question answering, document summarization, current information retrieval systems, etc. Later studies focused on temporal sentiment identification from social events. Fukuhara etal. [13] analyzed the temporal trends of sentiments and topics from a text archive that provided timestamps in weblogs and news

Pr oo

K16832_C013.indd 404

fs

fo rr

ev

ie w

on ly
9/5/2013 8:04:59 PM

Big Social Data Analysis

405

(a) 1 0.8 0.6 0.4 0.2 0 0.2 0.4

Sadness

Annoyed

Neutral Serene Final position

Absence of input information

0.2 0 0.2 0.4

Annoyed

Neutral Serene

Satised

Sadness

1.5 1 0.5 0 0.5

1 1.5

2 2.5

1.5 1 0.5 0 0.5 1 1.5 2 2.5

FIGURE 13.1 Ensemble sentiment streams obtained when fusing natural language data and facial expres sions, without (a) and with (b) Kalman filtering.

articles. The system produces two kinds of graphs: topic graph, which shows the temporal change of topics associated with a sentiment, and sentiment graph, which shows temporal change of sentiments associated with a topic. Mishne and de Rijke [28] proposed MoodViews, a system for the analysis of sentiment temporal changes based on LiveJournal mood tags. Similarly, Havre etal. [20] proposed ThemeRiver, a tool for visualizing thematic flows along with a timeline. In the context of multimodal sentiment analysis, finally, Cambria etal. [3] proposed a scalable methodology for fusing multiple cognitive and affective recognition modules in real time. Such a methodology is able to fuse any number of unimodal categorical modules, with very different time-scales and output labels. This is possible thanks to the use of a multidimensional vector space that provides the system with mathematical capabilities to deal with temporal issues. The proposed methodology outputs a continuous multidimensional stream that represents the users cognitive and affective progress over time in a novel and efficient way (Figure 13.1). A Kalman filtering technique controls the ensemble stream in real time to ensure temporal consistency and robustness. In particular, the methodology has been shown effective to fuse two different modalities: natural language data and facial expressions. In general, evaluation issues are largely solved for categorical recognition approaches. Unimodal categorical modules can be exhaustively evaluated, thanks to the use of large well-annotated databases and wellknown measures and methodologies (such as percentage of correctly classified instances, cross-validation, etc.). The evaluation of the performance of dimensional approaches is, however, an open and difficult issue to be solved. To this end, future works are expected to focus in depth on evaluation issues applicable to dimensional approaches and multimodality.

Pr oo

K16832_C013.indd 405

fs

fo rr

ev

ie w

on ly
Final position Satised 9/5/2013 8:05:00 PM

Ecstatic Disagreeable Enthusiastic Disgust Fear Initial Amused position Hopeful

Anger Panicky Attentive

(b)
Proceed Joy

1 0.8 0.6 0.4

Ecstatic Disagreeable Enthusiastic Initial Disgust Fear position Amused Hopeful

Anger Panicky Attentive

Proceed

Joy

406

Big Data Computing

Toward the Concept-Level Analysis of Big Sentiment Data

Standard approaches to opinion mining can be classified into three main categories: keyword spotting, lexical affinity, and statistical methods. All such approaches mainly rely on specific affect words that need to be explicitly used in the opinionated text. Hence, when emotions are conveyed implicitly, standard approaches are unable to correctly mine opinions and sentiments. More recent approaches to sentiment analysis, for example, the energy-based knowledge representation formalism [30] and sentic computing [4], in turn, work at concept-level in which they exploit semantic networks that allow one to take into account the conceptual and affective information implicitly conveyed by natural language concepts. Although scientific research in the area of emotion stretches back to the nineteenth century when Charles Darwin and William James proposed theories of emotion that continue to influence thinking today [8,22], the injection of affect into computer technologies is much more recent. During most of the last century, research on emotions was conducted by philosophers and psychologists, whose work was based on a small set of emotion theories that continue to underpin research in this area. The first researchers to try linking text to emotions were actually social psychologists and anthropologists who tried to find similarities on how people from different cultures communicate [31]. This research was also triggered by dissatisfaction with the dominant cognitive view centered around humans as information processors [26]. Later on, in the 1980s, researchers such as Turkle [38] began to speculate about how computers might be used to study emotions. Systematic research programs in this front began to emerge in the early 1990s. For example, Scherer [35] implemented a computational model of emotion as an expert system. A few years later, Picards landmark book affective computing [34] prompted a wave of interest among computer scientists and engineers looking for ways to improve humancomputer interfaces by coordinating emotions and cognition with task constraints and demands. Affect detection is critical because an affect-sensitive interface can never respond to users affective states if it cannot sense their affective states. Affect detection need not be perfect but must be approximately on target. Affect detection is, however, a very challenging problem because emotions are constructs (i.e., conceptual quantities that cannot be directly measured) with fuzzy boundaries and with substantial individual difference variations in expression and experience. To overcome such an issue, sentic computing* builds upon a brain-inspired and psychologically motivated affective categorization model [5] that can potentially describe the full range of emotional experiences in terms of four independent but concomitant dimensions, whose different levels of activation make up the total emotional state of the mind
*

Pr oo

http://sentic.net/sentics

K16832_C013.indd 406

fs

fo rr

ev

ie w

on ly
9/5/2013 8:05:00 PM

Big Social Data Analysis

407

Sensitivity Aptitude Attention Pleasantness

Vigilance Rage

no

ya

eness

0 1 e 1/18 2 2 1 e 4/18 2 2 1
B Disgust

nc

ie w
Pensiv

o ed or

Loathing

Sadness Surprise

Ac Sereni ty ce pt an ce

1 e4/18 2 2 1 e 1/18 2 2

Anger An

Anticipation

Grief

ev

Pleasantness

fo rr

Aptitude Sensitivity

FIGURE 13.2 The hourglass of emotions.

(Figure 13.2). In sentic computing, whose term derives from the Latin sentire (root of words such as sentiment and sentience) and sensus (intended both as capability of feeling and as common sense), the analysis of natural language is based on affective ontologies and common-sense reasoning tools, which enable the analysis of text not only at document-, page-, or paragraph-level, but also at sentence- and clause-level. In particular, sentic computing involves the use of AI and Semantic Web techniques, for knowledge representation and inference; mathematics, for carrying out tasks such as graph mining and multi-dimensionality reduction; linguistics, for discourse analysis and pragmatics; psychology, for cognitive and affective modeling; sociology, for understanding social network dynamics and social influence; finally ethics, for understanding related issues about the nature of mind and the creation of emotional machines. In order to effectively mine and analyze opinions and sentiments, it is necessary to bridge the gap between unstructured natural language data and structured machine-processable data. To this end, an intelligent software engine has been proposed, within the umbrella of sentic computing, to

Pr oo

K16832_C013.indd 407

fs

on ly
Ecstasy Joy Admiration Trust

est Inter

Ap p

re

he

Distra

ns

io

ction

Fear Terror

Amazement

Attention

9/5/2013 8:05:00 PM

408

Big Data Computing

Training corpus

CF-IOF

Spectral association Concept n-grams SBoCs Concept n-grams

IsaCore

Pre-processing
Text

lemmas

Sentic parser

AffectNet

Sentic medoids

extract the semantics and sentics (i.e., the conceptual and affective information) associated with natural language text, in a way that the opinions and sentiments in it contained can be more easily aggregated and interpreted. The engine exploits graph mining and multi-dimensionality reduction techniques on IsaCore [7] and AffectNet [4], respectively (Figure 13.3). Several other affect recognition and sentiment analysis systems [1,911,25,27,37] are based on different emotion categorization models, which generally comprise a relatively small set of categories (Table 13.1). The Hourglass of Emotions, in turn, allows the opinion-mining engine to classify affective information both in a categorical way (according to a wider number of emotion categories) and in a dimensional format (which facilitates comparison and aggregation). The opinion-mining engine consists of four main components: a preprocessing module, which performs a first skim of the opinion; the sentic parser, whose aim is to extract concepts from the opinionated text; the IsaCore module, for inferring the semantics associated with the given concepts; the AffectiveSpace module, for the extraction of sentics. The pre-processing module first exploits linguistic dictionaries to interpret all the affective valence indicators usually contained in opinionated text, for example, special punctuation, complete upper-case words, cross-linguistic onomatopoeias, exclamation words, degree adverbs, and emoticons. Second, the module detects negation and spreads it in such a way that it can be accordingly associated to concepts during the parsing phase. Handling

Pr oo

K16832_C013.indd 408

fs

fo rr

ev

FIGURE 13.3 Opinion mining engine block diagram. After performing a first skim of the input text, the engine extracts concepts from it and, hence, infers related semantics and sentics.

ie w

Hourglass model

Affective valence indicators

SBoCs

on ly
Truncated SVD

Semantics

Sentics

AffectiveSpace

9/5/2013 8:05:01 PM

Big Social Data Analysis

409

TABLE 13.1 An Overview of Recent Model-Based Sentiment Analysis Systems

Study [25] [10] [11] [9] [37] [27] [1] [15] [6] [7] Techniques Model Corpora Knowledge Base None ITS interaction ITS interaction ConceptNet WNA WNA None ConceptNet, WNA, HEO ConceptNet, WNA NB, SVM 2 categories Political articles LSA, MLP, NB, KNN 3 categories Dialog turns Cohesion indices 4 categories Dialog logs VSM, NB, SVM 5 categories ISEAR WN presence, LSA 6 categories News stories WN presence 6 categories Chat logs Winnow linear, C4.5 7 categories Children stories VSM, k-means 24 categories YouTube, LiveJournal VSM, neural 24 categories PatientOpinion networks VSM, k-medoids, 24 categories Twitter, LiveJournal, LDA PatientOpinion

Note: Studies are divided by techniques applied, number of categories of the model adopted, corpora, and knowledge base used.

negation is an important concern in opinion- and sentiment-related analysis, as it can reverse the meaning of a statement. Such task, however, is not trivial as not all appearances of explicit negation terms reverse the polarity of the enclosing sentence and that negation can often be expressed in rather subtle ways, for example, sarcasm and irony, which are quite difficult to detect. Lastly, the module converts text to lower-case and, after lemmatizing it, splits the opinion into single clauses according to grammatical conjunctions and punctuation. For parsing text, a modified version of COGPARSE [29] is exploited for quickly identifying meaningful stretches of text without requiring timeconsuming phrase structure analysis. The use of constructions, defined as stored pairings of form and function [17], makes it possible to link distributed linguistic components to one another, easing extraction of semantics from linguistic structures. Constructions are composed of fixed lexical items and category-based slots, or spaces that are filled in by lexical items during text processing. An interesting example from the relevant literature would be the construction [<ACTION> <OBJECT> <DIRECTION> <OBJECT>], instances of which include sneeze the napkin across the table or hit the ball over the fence. Constructions help us understand not only how various lexical items work together to create overall meanings, but also give the parser a sense of what categories of words are used together and thus which words to expect where. The sentic parser uses this knowledge to determine which constructions are matched by which lexical items and how good each match is. Each sentic parsers construction contributes its own unique semantics and carries a unique name. The sentic parser uses

Pr oo

K16832_C013.indd 409

fs

fo rr

ev

ie w

on ly
ConceptNet, Probase
9/5/2013 8:05:01 PM

410

Big Data Computing

knowledge about the lexical items found in text to choose the best possible construction for each span of text. Specifically, the parser looks each lexical item up in AffectNet and IsaCore, obtaining information about the basic category membership of that word. It then efficiently compares these potential memberships with the categories specified for each construction in the Corpus, finding the best matches so that, for example, the parser can extract a concept such as buy christmas present from a sentence such as today I bought a lot of very nice Christmas bells. Constructions are typically nested within one another: the sentic parser is capable of finding only those construction overlaps that are semantically sensible based on the overall semantics of constructions and construction slot categories, thus greatly reducing the time taken to process large numbers of texts. In the Big Data environment, a key benefit of construction-based parsing is that only small regions of correct text are required in order to extract meaning; word category information, and the generally small size of constructions mean that use can still be made of error-filled or otherwise normally unparseable texts. Texts do not have to have correct sentence-wide phrase structure in order to be parsed. Additionally, the sentic parser provides, for each retrieved concept, its relative frequency, valence, and status, that is, the concepts occurrence in the text, its positive or negative connotation, and the degree of intensity with which the concept is expressed, respectively. For each clause, the module outputs a small bag of concepts (SBoC), which is later on analyzed separately by the IsaCore and AffectiveSpace modules to infer the conceptual and affective information associated with the input text, respectively. In case any of the detected concepts is found more than once in the vector space (i.e., any of the concepts has multiple senses), all the SBoC concepts are exploited for a context-dependent coarse sense disambiguation. In particular, to represent the expected semantic value of the clause as a whole, the vectors corresponding to all concepts in the clause (in their ambiguous form) can be averaged together. The resulting vector does not represent a single meaning but the ad-hoc category of meanings that are similar to the various possible meanings of concepts in the clause [19]. Then, to assign the correct sense to the ambiguous concept, the concept sense with the highest dot product (and thus the strongest similarity) with the clause vector is searched. Once natural language text is deconstructed into concepts, these are given as input to both the Isacore and the AffectiveSpace modules. While the former exploits the graph representation of the common and common-sense knowledge base to detect semantics, the latter exploits the vector space representation of AffectNet to infer sentics. In particular, the IsaCore module applies spectral association for assigning activation to key nodes of the semantic network, which are used as seeds or centroids for classification. Such seeds can simply be the concepts corresponding to the class labels of interest plus their available synonyms and antonyms, if any. Seeds can

Pr oo

K16832_C013.indd 410

fs

fo rr

ev

ie w

on ly
9/5/2013 8:05:01 PM

Big Social Data Analysis

411

also be found by applying CF-IOF [4] on a training corpus (when available), in order to perform a classification that is more relevant to the data under analysis. After seed concepts are identified, the module spreads their values across the IsaCore graph. This operation, an approximation of many steps of spreading activation, transfers the most activation to concepts that are connected to the seed concepts by short paths or many different paths in affective common-sense knowledge. Therefore, the concepts of each SBoC provided by the sentic parser are projected on the matrix resulting from spectral association in order to calculate their semantic relatedness to each seed concept and, hence, their degree of belonging to each different class. Such classification measure is directly proportional to the degree of connectivity between the nodes representing the retrieved concepts and the seed concepts in the IsaCore graph. The concepts retrieved by the sentic parser are also given as input to the AffectiveSpace module, which, in turn, exploits dimensionality reduction techniques to infer the affective information associated with them. To this end, the concepts of each SBoC are projected into AffectiveSpace and, according to their position in the vector space representation of affective common-sense knowledge, they are assigned to an affective class defined through the sentic medoids technique [3]. As well as in the IsaCore module, the categorization does not consist in simply labeling each concept, but also in assigning a confidence score to each emotional label, which is directly proportional to the degree of belonging to a specific affective cluster (dot product between the given concept and the relative sentic medoid). Such affective information can also be exploited to calculate a polarity value associated with each SBoC provided by the sentic parser, as well as to detect the overall polarity associated with the opinionated text. In order for the engine to perform fast real-time opinion mining, the sentic vector associated with each AffectNet concept can be calculated a priori and saved in an SQL database. At runtime, then, the sentic vectors relative to each of the concepts composing the SBoC can be retrieved from such a database and aggregated to compute the overall affective information associated with the specific SBoC. This allows the sentic extraction process to be faster than directly applying the AffectiveSpace process. Similarly, spectral association can be computed a priori on IsaCore and the semantic classification of each concept (i.e., a set of different topic labels and the confidence associated with these) can be stored in an SQL database. On average, in fact, while the processing of a 100-word opinion is in the order of tens of seconds when directly applying the AffectiveSpace and IsaCore processes, the extraction of semantics and sentics is in the order of seconds when using the corresponding SQL databases. Both resources are in the order of hundreds of megabytes and, hence, easily exportable and embeddable into bigger systems for the development of applications in fields such as social media marketing, humancomputer interaction, and e-health [4].

Pr oo

K16832_C013.indd 411

fs

fo rr

ev

ie w

on ly
9/5/2013 8:05:01 PM

412

Big Data Computing

Conclusion
As society evolves at the same pace of the Web, online social data are becoming more and more important for both individuals and businesses. The Web 2.0, however, has unleashed an era of online participation that is causing user-generated contents to grow exponentially and, hence, to become continuously larger and more complex. In order to reduce such complexity, we need to adopt a multidisciplinary approach to big social data analysis, which combines different perspectives by means of social media analytics, trend discovery, multimedia management, social network analysis, opinion mining, and more. Specifically, in order to translate mere collected intelligence into true collective intelligence [18], we need to aim for a concept-level analysis of online social data that can enable the aggregation of the semantics and sentics associated with text. Next-generation opinion mining systems, thus, need broader and deeper affective common sense knowledge bases, coupled with more brain-inspired and psychologically motivated reasoning models, in order to better understand natural language opinions through more efficiently bridging the gap between (unstructured) human-processable information and (structured) machineprocessable data.

References

1. C. Alm, D. Roth, and R. Sproat. Emotions from text: Machine learning for textbased emotion prediction. In HLT/EMNLP, pp. 347354, Vancouver, 2005. 2. A. Brew, D. Greene, and P. Cunningham. Using crowdsourcing and active learning to track sentiment in online media. In ECAI, pp. 145150, Lisbon, 2010. 3. E. Cambria, N. Howard, J. Hsu, and A. Hussain. Sentic blending: Scalable multimodal fusion for the continuous interpretation of semantics and sentics. In IEEE SSCI, pp. 108117, Singapore, 2013. 4. E. Cambria and A. Hussain. Sentic Computing: Techniques, Tools, and Applications. Springer, Dordrecht, the Netherlands, 2012. 5. E. Cambria, A. Livingstone, and A. Hussain. The hourglass of emotions. In A. Esposito, A. Vinciarelli, R. Hoffmann, and V. Muller, editors, Cognitive Behavioral Systems, volume 7403 of Lecture Notes in Computer Science, pp. 144157. Springer, Berlin, 2012. 6. E. Cambria, T. Mazzocco, and A. Hussain. Application of multi-dimensional scaling and artificial neural networks for biologically inspired opinion mining. Biologically Inspired Cognitive Architectures, 4:4153, 2013 7. E. Cambria, Y. Song, H. Wang, and N. Howard. Semantic multi-dimensional scaling for open-domain sentiment analysis. IEEE Intelligent Systems, DOI 10.1109/MIS.2012.118, 2013.

Pr oo

K16832_C013.indd 412

fs

fo rr

ev

ie w

on ly
9/5/2013 8:05:01 PM

Big Social Data Analysis

413

8. C. Darwin. The Expression of the Emotions in Man and Animals. John Murray, London, 1872. 9. T. Danisman and A. Alpkocak. Feeler: Emotion classification of text using vector space model. In AISB, Aberdeen, 2008. 10. S. DMello, S. Craig, J. Sullins, and A. Graesser. Predicting affective states expressed through an emote-aloud procedure from autotutors mixed-initiative dialogue. International Journal of Artificial Intelligence in Education, 16:328, 2006. 11. S. DMello, N. Dowell, and A. Graesser. Cohesion relationships in tutorial dialogue as predictors of affective states. In Conf. Artificial Intelligence in Education, pp. 916, Brighton, 2009. 12. U. Fayyad, G. Piatetsky, and P. Smyth. From data mining to knowledge discovery: An overview. In Advances in Knowledge Discovery and Data Mining, U. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy, eds. pp. 136. MIT Press, Cambridge, 1996. 13. T. Fukuhara, H. Nakagawa, and T. Nishida. Understanding sentiment of people from news articles: Temporal sentiment analysis of social events. In ICWSM, Boulder, 2007. 14. X. Glorot, A. Bordes, and Y. Bengio. Domain adaptation for large-scale sentiment classification: A deep learning approach. In ICML, Bellevue, 2011. 15. M. Grassi, E. Cambria, A. Hussain, and F. Piazza. Sentic web: A new paradigm for managing social media affective information. Cognitive Computation, 3(3):480489, 2011. 16. N. Godbole, M. Srinivasaiah, and S. Skiena. Large-scale sentiment analysis for news and blogs. In ICWSM, Boulder, 2007. 17. A. Goldberg. Constructions: A new theoretical approach to language. Trends in Cognitive Sciences, 7(5):219224, 2003. 18. T. Gruber. Collective knowledge systems: Where the social web meets the semantic web. Web Semantics: Science, Services and Agents on the World Wide Web, 6(1):413, 2007. 19. C. Havasi, R. Speer, and J. Pustejovsky. Coarse word-sense disambiguation using common sense. In AAAI CSK, Arlington, 2010. 20. S. Havre, E. Hetzler, P. Whitney, and L. Nowell. Themeriver: Visualizing thematic changes in large document collections. IEEE Transactions on Visualization and Computer Graphics, 8(1):920, 2002. 21. M. Hearst. Text data mining: Issues, techniques, and the relationship to information access. In UW/MS Workshop on Data Mining, Berkeley, 1997. 22. W. James. What is an emotion? Mind, 34:188205, 1884. 23. V. Khuc, C. Shivade, R. Ramnath, and J. Ramanathan. Towards building largescale distributed systems for twitter sentiment analysis. In ACM Symposium on Applied Computing, pp. 459464, Riva del Garda, 2012. 24. O. Kucuktunc, B. Cambazoglu, I. Weber, and H. Ferhatosmanoglu. A large-scale sentiment analysis for Yahoo! answers. In ACM International Conference on Web Search and Data Mining, pp. 633642, Seattle, 2012. 25. W. Lin, T. Wilson, J. Wiebe, and A. Hauptmann. Which side are you on? Identifying perspectives at the document and sentence levels. In Conference on Natural Language Learning, pp. 109116, Boston, 2006. 26. C. Lutz and G. White. The anthropology of emotions. Annual Review of Anthropology, 15:405436, 1986.

Pr oo

K16832_C013.indd 413

fs

fo rr

ev

ie w

on ly
9/5/2013 8:05:01 PM

414

Big Data Computing

Pr oo

27. C. Ma, A. Osherenko, H. Prendinger, and M. Ishizuka. A chat system based on emotion estimation from text and embodied conversational messengers. In International Conference on Active Media Technology, pp. 546548, Kagawa, 2005. 28. G. Mishne and M. de Rijke. Capturing global mood levels using blog posts. In AAAI, pp. 145152, Boston, 2006. 29. D. Olsher. COGPARSE: Brain-inspired knowledge-driven full semantics parsing. In Advances in Brain Inspired Cognitive Systems, Volume of 7366 Lecture Notes in Computer Science, pp. 111, 2012. 30. D. Olsher. Full spectrum opinion mining: Integrating domain, syntactic and lexical knowledge. In ICDM, Brussels, 2012. 31. C. Osgood, W. May, and M. Miron. Cross-Cultural Universals of Affective Meaning. University of Illinois Press, Brussels, pp. 693700, 1975. 32. B. Pang and L. Lee. Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2:1135, 2008. 33. B. Pang, L. Lee, and S. Vaithyanathan. Thumbs up? Sentiment classification using machine learning techniques. In EMNLP, pp. 7986, Philadelphia, 2002. 34. R. Picard. Affective Computing. The MIT Press, Boston, 1997. 35. K. Scherer. Studying the emotion-antecedent appraisal process: An expert system approach. Cognition and Emotion, 7:325355, 1993. 36. E. Simoudis. Reality check for data mining. IEEE Expert, 11(5), 1996. 37. C. Strapparava and R. Mihalcea. Learning to identify emotions in text. In ACM Symp. Applied Computing, pp. 15561560, Fortaleza, 2008. 38. S. Turkle. The Second Self: Computers and the Human Spirit. Simon & Schuster, New York, 1984. 39. P. Turney. Thumbs up or thumbs down? Semantic orientation applied to unsupervised classification of reviews. In ACL, pp. 417424, Philadelphia, 2002. 40. C. Wayne. Multilingual topic detection and tracking: Successful research enabled by corpora and evaluation. In LREC, pp. 14871494, Athens, 2000. 41. Y. Wilks and M. Stevenson. The grammar of sense: Using part-of-speech tags as a first step in semantic disambiguation. Journal of Natural Language Engineering, 4(2):135143, 1998. 42. H. Wu, R. Luk, K. Wong, and K. Kwok. Interpreting TF-IDF term weights as making relevance decisions. ACM Transactions on Information Systems, 26(3), Article 13, 2008.

K16832_C013.indd 414

fs

fo rr

ev

ie w

on ly
9/5/2013 8:05:02 PM