Cognitive Computation https://doi.org/10.1007/s12559-021-09824-x Ten Years of Sentic Computing Yosephine Susanto1 · Erik Cambria1 · Bee Chin Ng1 · Amir Hussain2 Received: 30 December 2020 / Accepted: 6 January 2021 © The Author(s), under exclusive licence to Springer Science+Business Media, LLC part of Springer Nature 2021 Abstract Sentic computing is a multi-disciplinary approach to sentiment analysis at the crossroads between afective computing and commonsense computing, which exploits both computer and social sciences to better recognize, interpret, and process opin- ions and sentiments over the Web. In the last ten years, many diferent models (such as the Hourglass of Emotions and Sentic Patterns), resources (such as AfectiveSpace and SenticNet), algorithms (such as Sentic LDA and Sentic LSTM), and applica- tions (such as Sentic PROMs and Sentic Album) have been developed under the umbrella of sentic computing. In this paper, we review all such models, resources, algorithms, and applications together with the key shifts and tasks introduced by sentic computing in the context of afective computing and sentiment analysis. We also discuss future directions in these felds. Introduction only at document, page, or paragraph level, but also at sen- tence, clause, and concept level. With the recent development of deep learning, research in In this paper, we review key sentic computing models, artifcial intelligence (AI) has gained new vigor and promi- resources, algorithms, and applications together with the nence. Machine learning, however, sufers from three big works that have been using them in the context of afective issues, namely: computing and sentiment analysis during the last decade. The remainder of this paper is organized as follows: Sen- 1. Dependency: it requires (a lot of) training data and is tic Computing’s Key Shifts describes the three key shifts domain dependent; introduced by sentic computing; Sentic Computing’s Key 2. Consistency: diferent training or tweaking leads to dif- Tasks lists the ffteen key tasks of sentic computing; Sen- ferent results; tic Computing’s Key Models illustrates the two key models 3. Transparency: the reasoning process is uninterpretable on which sentic computing is based; Sentic Computing’s (black-box algorithms). Key Resources introduces two key sentic resources; Sentic Computing’s Key Algorithms explains two key sentic algo- Sentic computing [1] addresses such issues in the context rithms; Sentic Computing’s Key Applications showcases of natural language processing (NLP) through a multi- two key sentic applications; Future Directions discusses disciplinary approach that aims to bridge the gap between future directions; fnally, Conclusion provides concluding statistical NLP and many other disciplines that are neces- remarks. sary for understanding human language, such as linguistics, commonsense reasoning, semiotics, and afective comput- ing. Sentic computing, whose term derives from the Latin Sentic Computing’s Key Shifts sensus (as in commonsense) and sentire (root of words such as sentiment and sentience), enables the analysis of text not Sentic computing’s new approach to NLP gravitates around three key shifts: 1. Shift from mono- to multi-disciplinarity—evidenced * Erik Cambria by the concomitant use of symbolic and subsymbolic AI, [email protected] for knowledge representation and reasoning; semiotics, for 1 School of Computer Science and Engineering, Nanyang meaning encoding and decoding; mathematics, for carry- Technological University, Nanyang, Singapore ing out tasks such as graph mining and multidimensionality 2 School of Computing, Edinburgh Napier University, reduction; linguistics, for discourse analysis and pragmatics; Edinburgh, United Kingdom Vol.:(0123456789)1 3 Cognitive Computation concept level entails preserving the meaning carried by mul- tiword expressions such as cloud_computing, which repre- sent ‘semantic atoms’ that should never be broken down into single words. In the bag-of-words model, for example, the concept cloud_computing would be split into computing and cloud, which may wrongly activate concepts related to the weather and, hence, compromise categorization accuracy. 3. Shift from statistics to linguistics—implemented by allowing sentiments to fow from concept to concept based on the dependency relation between clauses (Fig. 3). The sentence “iPhone12 is expensive but nice”, for example, is equal to “iPhone12 is nice but expensive” from a bag-of- words perspective. However, the two sentences bear oppo- site polarity: the former is positive as the user seems to be willing to make the efort to buy the product despite its high price, and the latter is negative as the user complains about the price of iPhone12 although he/she likes it. Fig. 1 Sentic computing disciplines Sentic Computing’s Key Tasks Sentic computing takes a holistic approach to natural language psychology, for cognitive and afective modeling; sociology, understanding by handling the many sub-problems involved in for understanding social network dynamics and social infu- extracting meaning and polarity from text. While most works ence; fnally ethics, for understanding-related issues about approach it as a simple categorization problem, in fact, sen- the nature of mind and the creation of emotional machines timent analysis is actually a suitcase research problem that (Fig. 1). requires tackling many NLP tasks (Fig. 4). As Marvin Min- 2. Shift from syntax to semantics—enabled by the adop- sky would say, the expression ‘sentiment analysis’ itself is a tion of the bag-of-concepts model instead of simply counting big suitcase (like many others related to afective computing, word co-occurrence frequencies in text (Fig. 2). Working at e.g., emotion recognition or opinion mining) that all of us use Fig. 2 Jumping NLP curves [2] 1 3 Cognitive Computation personality recognition [7], for distinguishing between difer- ent personality types of the users, and more. Such structure is inspired by the jumping NLP curves paradigm (Fig. 2) and con- sists of 15 NLP tasks organized into three layers: 1 Syntactics layer—which aims to preprocess text so that informal text is reduced to plain English, infected forms of verbs and nouns are normalized, and basic sentence structure is made explicit. 2 Semantics layer—which aims to deconstruct the nor- malized text obtained from the syntactics layer into concepts, resolve references (that is, named entities and anaphora), and flter out neutral content from the input to improve sentiment classifcation accuracy. 3 Pragmatics layer—which aims to extract meaning from both sentence structure and semantics obtained from syn- tactics and semantics layers, respectively. After performing some kind of user profling (personality and sarcasm detec- Fig. 3 Sentic computing framework [1] tion), the pragmatics layer interprets metaphors (if any) and extracts opinion targets and the polarity associated with each of them. to encapsulate our jumbled idea about how our minds convey emotions and opinions through natural language. Sentic computing addresses the composite nature of the prob- Sentic Computing’s Key Models lem via a three-layer structure that concomitantly handles tasks such as microtext normalization [4], to decode informal text, The symbolic part of the sentic computing engine leverages subjectivity detection [5], to flter out neutral content, anaphora two key models, which regulate how emotions are assigned to resolution [6], to link pronouns with the entities of a sentence, specifc words and multiword expressions in a sentence and Fig. 4 Sentiment analysis is a suitcase problem [3] 1 3 Cognitive Computation Fig. 5 The Hourglass of Emo- tions [22] how such emotions fow throughout the sentence to determine Hourglass Model its polarity, respectively. This section describes these two models, namely the Hourglass of Emotions, a brain-inspired The Hourglass of Emotions is a new emotion model that and psychologically motivated emotion categorization model goes beyond mere categorical and dimensional approaches (Hourglass Model), and Sentic Patterns, sentiment-specifc lin- (Fig. 5). Beside emotion classifcation, the model has been guistic patterns that model how polarity fows from concept used for tasks like polarity detection from text [8, 9], audio to concept based on the dependency tree of sentences (Sentic and video [10, 11], and multiple languages [12], but also Patterns). knowledge representation [13], psycholinguistics [14], 1 3 Cognitive Computation Fig. 6 Emotion classifcation with fve sample emotion words for each category cognitive and cultural modeling [15, 16], social network activation of diferent emotional confgurations, resembling analysis [17], and the arts [18–21]. The Hourglass model Minsky’s k-lines [23]. represents afective states both through labels and through The model, in fact, is based on the idea that the mind is four independent but concomitant affective dimensions, made of diferent independent resources and that emotional namely Introspection (the joy-versus-sadness dimension), states result from turning some set of these resources on and Temper (the calmness-versus-anger dimension), Attitude (the turning another set of them of [24]. Each such selection pleasantness-versus-disgust dimension), and Sensitivity (the changes how we think by changing our brain’s activities: eagerness-versus-fear dimension). the state of anger, for example, appears to select a set of Each afective dimension is characterized by six levels of resources that help