The Nature of Emotional Expression in Social Media: Measurement, Inference and Utility

Transcription

1 The Nature of Emotional Expression in Social Media: Measurement, Inference and Utility Munmun De Choudhury Microsoft Research Redmond, WA ABSTRACT Today, social media platforms like Twitter and Facebook enable in-the-moment reflection of people's attitudes, attention, and emotions in a scale that was never available before. In this paper, we present and explore how we can measure, infer and utilize expression of human moods from social media activity (e.g., Twitter). Motivated by literature in psychology, we first measure more than 200 nuanced human moods at scale on Twitter, through a popular representation, known as the circumplex model that characterizes affective experience through two dimensions: valence and activation. Second, moving beyond explicit mood signals, we develop an automated classifier to infer several human affective states in social media. Starting with moods, we derive naturalistic signals from Twitter posts about a variety of affects of individuals and deploy them in a classification framework with promising results. Finally, we illustrate the utility of emotion exploration in social media via a case study that tracks behavioral change of new mothers post child-birth. Our findings provide at-scale naturalistic assessments and extensions of existing conceptualizations of human mood, as well as indicate their utility in domains of societal interest, such as public health. Author Keywords Affect, circumplex model, classification, emotion, mood, social media, Twitter ACM Classification Keywords H.5.m [Information Systems]: Information Systems Applications Miscellaneous. General Terms Algorithms; Human Factors; Measurement. INTRODUCTION The dynamic nature of user-generated content today holds formidable power in a culture that is increasingly shaped and influenced by technology. In particular, the multitude of various social media and social networking sites has Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Scott Counts Microsoft Research Redmond, WA counts@microsoft.com provided easy channels of opinion and emotional expression to large audiences. On top of such ubiquity, since social media sites provide a conducive platform to provide and receive social support around issues surrounding our daily lives, such in-the-moment emotional expression of current experiences has been on the rise in the past few years 1 : be it about the global economy broadly, the riots in Egypt, or a personal vacation trip with one s family. As a timely research focus, judicious tracking and monitoring of emotion at scale could potentially trigger a societal shift with enough support and publicity. Furthermore, they can enable new information-seeking approaches; for instance, identifying search features given an emotion attribute, or enabling effective emotionallyreflective interfaces. Consequently, there is significant value to be derived from measuring and classifying/inferring human emotion in social media that can potentially be utilized to impact several domains of public interest including health, finance, entertainment, advertising, politics or evolution of language and culture. What are emotions? According to literature in psychology, emotions are complex patterns of cognitive processes, physiological arousal, and behavioral reactions [11]. Emotions arouse us to action and direct and sustain that action. They also help us organize our experience by directing attention, and by influencing our perceptions of self, others, and our interpretation and memory of events [21]. As renowned sociologist Herbert Blumer postulated, what goes on around us sinks into the reservoirs of our mind and changes how we think. Given the critical value in understanding human emotion, researchers have attempted to describe emotion (alternatively affect) through a set of dimensions, typically Positive Affect and Negative Affect, where the two dimensions vary independently of each other [10,12]. However research in psychology provides evidence that rather than being independent, these dimensions are interrelated in a highly systematic fashion [13]. Consequently, psychology researchers devised a psychometric instrument called the circumplex [20], a 1

2 spatial model in which affective concepts fall on a circle and enable quantification of the cognitive structure of affective experience. The circumplex represents each affective concept (or, mood) via two dimensions: a pleasure-displeasure measure, known as valence, and a degree-of-arousal dimension, known as activation. Motivated along these lines, in this work, we utilize the circumplex model as a tool to measure more than 200 finegrained human moods at scale, as expressed on social media (Twitter). We also analyze several attributes of collective behavior in the light of measurement of moods: including usage levels, linguistic diversity as well as activity rates associated with different mood expression on social media. We thereby reveal how we can provide atscale naturalistic assessments and extensions of existing conceptualizations of moods (extended version in [4]). However notice that although measuring these moods in the light of human behavior can lend us useful insights, they do not help us infer moods broadly in the scope of any arbitrary social media content. For example, the discussion presented so far (lexicon-driven approach) relies on the explicit presence of a mood in a Twitter post; however in most cases, statistically, this might be a small percentage. Hence an automated method of inferring the broad mood (or affective state) might yield better coverage. Automatic classification of affect might be useful in domains where understanding the expressed mood of an author has some practical utility: that could range from advertising, recommendations, to tracking behavior of populations in health, socio-economics or urban development domains. Hence we utilize a wide range of moods to act as supervisory ground truth signals in the development of an affect classifier (see [5] for details). The classifier can infer a range of fine-grained affective states from posts on social media (beyond simply the valence of the affect descriptors), that might not bear explicit signals about emotion. Such inferred states would reflect the specific content, language and state of the individual sharing the content, i.e., the distinctive qualities of individuals affects. Inevitably, there is considerable practical utility in measuring and inferring human emotion at scale. For instance, we rely on the general observation that, emotionbearing content, such as dark (negative) postings can serve as signs of depression of individuals and can be utilized as an early warning system for timely intervention or to reduce non-invasively, the stigma around mental illness. Motivated along these lines, in the final part of the paper, we explore a particular utility domain of interest: investigating behavioral and emotional change of new mothers in the postnatal phase, compared to their prenatal behavior, based on their postings on Twitter. The rest of this paper is organized as follows. Following a review of prior work, we discuss the three major segments of interest in the paper in three sections: measuring moods, automatic classification of affect, and finally a case study to illustrate the utility of studying emotion dynamics in the context of social media. We conclude with a discussion of our major findings, and future research directions. BACKGROUND LITERATURE Considerable research in psychology has defined and examined human emotion and mood (e.g., [8]), with basic moods encompassing positive experiences like joy and acceptance, negative experiences like anger and disgust, and others like anticipation that are less clearly positive or negative. The activation attribute of moods, together with valence (i.e., the degree of positivity/negativity of a mood) characterize the structure of affective experience (ref. PAD emotional state model in [13]; also the circumplex model of affect in [20]). In these works, authors utilized self-reports of affective concepts to scale and order emotion types on the pleasure-displeasure scale (valence) and the degree-ofarousal scale (activation) based on perceived similarity among the terms. Turning to analyses of mood in social media, early work focused on sentiment in weblogs. Mihalcea et al. [14] utilized happy/sad labeled blog posts on LiveJournal to determine temporal and semantic trends of happiness. Similarly, Mishne et al. [16] utilized LiveJournal data to build models that predict levels of various moods and understand their seasonal cycles according to the language used by bloggers. More recently, Bollen et al. [1] analyzed trends of public moods (using a psychometric instrument POMS to extract six mood states) in light of a variety of social, political and economic events. Research involving affect exploration on social websites such as Facebook and Twitter has looked at trends of use of positive and negative words [12]. Recently, Golder et al. [10] studied how individual mood varies from hour-to-hour, day-to-day, and across seasons and cultures by measuring positive and negative affect in Twitter posts, using the lexicon LIWC ( Prior research has also tackled automatic classification of sentiment in online domains [19]. These machine learning techniques need extensive manual labeling of text for creating ground truth. Some of these issues have been tackled by utilizing emoticons present in text as labels for sentiment [3], although they tend to perform well mostly in the context of the two basic positive and negative affect classes. The closest attempt towards multiclass affect classification has been on LiveJournal data: the mood tags associated with posts were used as ground truth [15]. An alternative that circumvents the problems of machine learning techniques has been the use of generic sentiment lexicons such as WordNet, LIWC, and other lists [9]. Recently, there has been a growing interest in crowdsourcing techniques to manually rate polarity in Twitter posts [16].

3 Limitations Despite widespread interest, it is clear that the notions of affect and sentiment have been rather simplified in current state-of-the-art, often confined to their valence measure (positive/negative), with the six moods in [1] being an exception. However, as indicated, the psychology literature suggests that understanding the inter-relatedness of valence and activation is important in conceptualizing human emotion. Additionally, when it comes to classifying or inferring emotional attributes, hand-labeled ground truth for affect classification or manually curated word lists are likely to be unreliable and scale poorly on noisy, topicallydiverse social media data, such as Twitter. Finally, another problem with polarity-centric sentiment classifiers is that they typically encompass a vague notion of polarity that includes emotion, and opinion; and lumps them all into two classes positive and negative, refer [23]. In order to better make sense of emotional behavior on social media, we require a principled notion of human emotion a contribution discussed in this paper. MEASUREMENT OF HUMAN MOODS We will begin with our methodology of measuring emotion, in terms of moods, and understanding behavioral differences found in mood expression across individuals at scale. As pointed out earlier, we note that despite considerable interest in mining and analyzing human emotion, simplification of emotions to merely positive and negative dimensions may miss important nuances in mood expression. For instance, annoyed and frustrated are both negative, but they express two very different emotional states. A primary research challenge, therefore, is finding a principled way to identify a set of words that can truly help us measure emotional states of individuals, as well as characterize their nuances in terms of both their valence and arousal measure, known as activation. Hence we study a popular representation of human mood landscape: the circumplex model. For the purpose, we first present a systematic method to identify moods in social media that captures the broad range of individuals emotional states using mechanical turk studies and forays into the psychology literature. Second we perform analysis of these moods in the light of individuals behavioral attributes to reveal nuances of collective mood expression. Construction of Mood Lexicon We begin by discussing our methodology to construct a mood lexicon a set of words that would indicate individuals broad emotional states. We then characterize the moods by the two dimensions of valence and activation, and discuss how we infer them. We utilize five established sources to develop a mood lexicon: 1. ANEW: ANEW (Affective Norms for English Words) that provides a set of normative emotional ratings for ~2000 words in English, including their valence, activation and dominance ratings [2]. 2. LIWC: For LIWC, we use sentiment-indicative categories like positive/negative emotions, anxiety, anger and sad ( 3. EARL: Emotion Annotation and Representation Language dataset that classifies 48 emotions ( 4. A list of basic emotions provided in [18], e.g., fear, contentment, disgust etc. 5. A list of moods provided by the blogging website LiveJournal ( However, this large ensemble of words is likely to contain several words that do not necessarily define a mood (e.g., sleepy is a state of a person, rather than a mood). To circumvent this issue, we first perform a mechanical turk study ( to narrow our candidate set to truly mood-indicative words. In our task, each word had a 1 7 Likert scale (1 indicated not a mood at all, 7 meant definitely a mood). Only turkers from the U.S. and having an approval rating greater than 95% were allowed. Combining 12 different turkers ratings, we construct a list of those words where the median rating was at least 4, and the standard deviation was less than or equal to 1. The final set of mood words contained 203 terms (examples include: excited, nervous, quiet, grumpy, depressed, patient, thankful, bored). Given the final list of representative moods, our next task was to determine the values of the valence and activation dimensions of each mood. For those words in the final list that were present in the ANEW lexicon, we use the sourceprovided measures of valence and activation, as these values in the ANEW corpus had already been computed after extensive and rigorous psychometric studies. For the remaining words, we conduct another turk study, to systematically collect these measurements. Like before, we considered only those turkers who had at least 95% approval rating history and were from U.S. For a given mood, each turker (12 turkers for each word) was asked to rate the valence and activation measures, on two different 1 10 Likert scales (1 indicated low valence/low activation, while 10 indicated high values). Finally, we used the mean ratings for each word as the final measures its valence and activation (Fleiss-Kappa measure was 0.65). Data For data collection, we utilized the Twitter Firehose and focused on one year's worth of Twitter posts posted in English from Nov 1, 2010 to Oct 31, From this ensemble, in order to obtain Twitter posts for each of the 203 moods, we resorted to a method that could infer moodcontaining posts reasonably consistently and in a principled manner. We conjecture that posts containing moods as hashtags at the end are likely to capture the emotional state of the individual, in the limited context of the post. This is motivated by prior work where Twitter's hashtags and smileys were used as labels for sentiment classifiers [3]. We

4 Q2 Q3 Figure 2. Circumplex model showing usage frequencies of moods used as hashtags at the end of Twitter posts: larger squares represent higher frequency of usage. also referred to the study in [4] which indicated that that in 83% of the cases, hashtagged moods at the end of posts indeed captured the users' moods. For instance, #iphone4 is going to be available on verizon soon! #excited expresses the mood excited. By this process, our labeled mood dataset comprised about 10.6M tweets from about 4.1M users. Deciphering Human Behavior through Moods We now study various aspects of the relationship between mood expression and human behavior in social media (Twitter). We present four aspects of such behavior: usage levels, diurnal patterns, diversity of language use, and activity rate, given a mood, in the next three subsections. Usage Levels of Moods Our study of mood exploration on Twitter data is based on analyzing the circumplex model of moods in terms of the moods usage frequencies. We illustrate these mood usage frequencies (count over all posts) on the circumplex model in Figure 2, where the size of squares (i.e., moods) is proportional to its frequency. We note that the usages of moods in each of the quadrants is considerably different (the differences between each pair of quadrants were found to be statistically significant based on independent sample t- tests: p<0.0001). The overall trend shows that moods in Q3 (low valence, low activation) tend to be used extensively (sad, bored, annoyed, lazy), along with a small number of moods in Q1, of relatively higher valence and activation (happy, optimistic). Overall, usage frequencies of lower valence moods exceed those of higher valence moods. Q4 We hypothesize the presence of a broadcasting bias behind these observations. Since individuals often use Twitter to broadcast their opinions and feelings on various topics, it is likely that the mood about some information needs to be of sufficiently low or high valence to be worth reporting to the audience. This appears to be particularly true with respect to positive valence terms, with mildly positive moods expressed only rarely. The observation that lower valence moods are shared more often might be due to individuals seeking social support from their audiences in response to various happenings externally as well as in their own lives. In a sense, via observing the usage levels of moods, we are able to validate the topology of the circumplex model in the social media context and showed that all moods are after all not created equal! Temporal Patterns of Mood Use We follow with the trail of investigating mood usage in the light of diurnal and weekly behavioral patterns. In Figure 1 we show the usage of the 203 moods in Twitter posts over a 24 hour period, averaged over all users and all days in our dataset (data points within the 95% confidence interval are considered). We divide usage over the course of a day into four different timings: morning (5am-12pm), afternoon (12pm-5pm), evening (4pm-10pm), nightowl (10pm-5am). The main observation from the circumplex model of the diurnal patterns is that greater usage activity in general, in terms of mood expression occurs during the evening and night; with morning have the least usage. Evenings show Figure 1. Circumplex model showing mood use diurnally over a 24 hour period. Each day is divided into four types of usage timings: morning (5am-12pm), afternoon (12pm-5pm), evening (4pm-10pm) and nightowl (10pm-5am).

5 Q2 Q1 Q1 Figure 4. Circumplex model showing mood use during weekdays (Monday-Thursday); weekends (Friday-Sunday). Weekdays show more negative, higher activation moods. both high use of negative and positive valence moods (e.g., negative moods: unhappy, tired, sad, lazy, worthless; positive moods: win, nice, good, terrific); while at nights, the overall activation of moods used tends to increase compared to other times of the day (e.g., terrified, blocked, stressed). Intuitively this appears to conform to common expectations: people are likely to feel tired and worthless at the end of a work day; at the same time certain others probably feel blocked and stressed at nights (also ref. [10]). We further present temporal patterns of mood usage during weekdays (Monday-Thursday) and weekends (Friday-Sunday) in Figure 4. The primary finding from the two circumplex models is that positive moods are more extensively used during weekends, compared to weekdays; whereas negative ones are predominant during the weekdays. Additionally a bulk of the moods shared on weekdays has higher activation than others. This probably is reflective of people being busy around a work week, while tending to be more relaxed during the weekends, thereby showing lower activation in mood expression. Linguistic Diversity of Moods We now explore how the diversity of linguistic content associated with usage of various moods relates to their valence and activation. Like before, we show the moods (as squares) on the circumplex model (Figure 3). The color of the squares indicate a mood s normalized entropy, defined as the entropy of the textual content (i.e., unigrams over all posts associated with the mood), divided by the total number of posts expressing the mood. In the figure, lighter shades indicate higher entropy. We observe that squares on the right side of the circumplex model (Q1, Q4) tend to have higher entropy than left (Q2, Q3) (statistically significant based on an independent sample t-test). This indicates that while positive moods tend to be shared across a wide array of linguistic context (topics, events etc.), negative moods tend to be shared in a limited context, confined to limited topics. Activity Rates of Moods Next we investigate the relationship between an individual s activity rate and his/her mood expression. We Q3 Figure 3. Circumplex model showing entropies of moods in terms of the content of posts: higher valence moods (shown in lighter shades) have higher entropy. define a measurement of how active s/he is in sharing posts: the number of posts shared per second since the time of the individual s account creation. Based on this definition, we show the circumplex model of moods in Figure 5. The size of each square in this case is proportional to the mean rate of activity of all individuals who have shared the particular mood. The figure shows that the majority of the larger squares (or moods shared by highly active individuals) lie in Q1 and Q4; in other words, high (or positive) valence moods are shared by highly active individuals (statistically significant based on independent sample t-tests between quadrant pairs). On the other hand, moods of high activation (in Q2) but low valence are shared primarily by individuals with a low activity rate. In general, this indicates that positive moods are associated more frequently with active individuals, while negative and high arousal moods appear to be shared more frequently by individuals with low activity rates. Through the studies so far, our major contribution thus lies in the large-scale naturalistic validation of the circumplex model encompassing a variety of human moods expressed online. However, we note that, on noisy and immensely large and diverse streams like social media, collecting explicit signals on moods (e.g. in the form of hashtags) might pose as a significant research challenge [4, 5]. In the following section, we therefore present an automatic classification framework that can infer affective states automatically. AUTOMATED INFERENCE OF AFFECTIVE STATES We discuss how the 203 moods can be utilized to infer a wide variety of human affective states in social media in an automated manner. The primary challenge in classifying Q4

6 Q2 Q3 affect lies in the unavailability of ground truth: an aspect often circumvented via manual labeling in order to create training examples. As we move to social media domains featuring enormous data, coupled with unavailability of ground truth, gathering appropriate training data necessitates a scalable alternative approach. Besides, in a typical sentiment classification setting, two broad, general factors Positive Affect (PA) and Negative Affect (NA) have emerged reliably as the dominant dimensions of emotional experience. However it is imperative to account for more distinguishable and fine-grained affective states. In this section we propose an affect classifier of social media data, along with promising results, that does not rely on any hand-built list of features or words, except for the near 200 mood hashtags that we use as a supervision ground truth signal (extended results in [5]). Inferring Affective States from Moods The psychology literature indicates that there is an implicit relationship between the externally observed affect and the internal mood of an individual [22]. When affect is detected by an individual (e.g., smile as an expression of joviality), it is characterized as an emotion or mood. In the subsection, we, therefore, discuss how we can utilize our previously discussed mood lexicon to find a mapping to broad affective states. Affect Types. Although affect has been found to comprise broadly positive and negative dimensions (PANAS positive and negative affect schedule [22]), we are interested in more fine-grained representation of human affect. Hence we utilize a source known as PANAS-X [22]. PANAS-X defines 11 specific affects apart from the positive and negative dimensions fear, sadness, guilt, Q1 Q4 Figure 5. Circumplex model of moods showing the relationship between mood expression and activity (twitter posts made per second by an individual sharing the mood). Larger squares indicate higher activity. hostility, joviality, self-assurance, attentiveness, shyness, fatigue, surprise, and serenity. Affect #moods Affect #moods Joviality 30 Fear 14 Fatigue 19 Guilt 5 Hostility 17 Surprise 8 Sadness 38 Shyness 7 Serenity 12 Attentiveness 2 Self-assurance 20 Table 1. Number of moods associated with the affects. Mood to Affect Associations. Next, based on the mapping of moods to affects provided in the PANAS-X literature, we derived associations for 60% moods from our lexicon. For the remaining associations, we conducted a turk study [5]. Each turker (12 in all, per mood) was asked to select the most appropriate affect that described a particular mood. We combined the ratings per mood, and used the affect that received majority rating (Fleiss-Kappa was 0.7). The distribution of #moods over affects is shown in Table 1. Training Data Collection Like before, we utilized Twitter Firehose data (English language posts between Nov 1, 2010 and Oct 31, 2011) for constructing labeled training examples for affect classification. We begin with our mood labeled dataset (ref. previous section) and then utilize the mood-affect mapping to labels posts corresponding to each affect. For instance, #iphone4 is going to be available on verizon soon! #excited expresses the mood excited, which can subsequently be mapped to the affect joviality. We further eliminated RT (retweet) posts to capture true personal affect reflection in a post. Classification and Experimental Results We use a classification setup that is standard in text classification as well as in sentiment classification. We represent Twitter posts as vectors of unigram and bigram features. Before feature extraction, the posts are lowercased, numbers are normalized into a canonical form, and URLs are removed. Finally the posts are tokenized. After feature extraction, features that occur fewer than five times are removed in a first step of feature reduction. We then randomly split the data into three folds for crossvalidation. The classification algorithm is a standard maximum entropy classifier. For each fold, we deploy this classifier to predict the affect labels of the test portion of the fold (33.3%), after training on the training portion (66.6%) of the fold. We begin by discussing the performance of classifying the Twitter posts in our dataset into the 11 different affect classes. We report the mean F1 measures for the 11 affect classes in Table 2. Our results show that the performance (precision/recall) of various affect classes differs widely. Noting the mood distributions for the various affects in Table 1, it appears

7 from the F1 measures in Table 2, that the good performance for the affects joviality, fatigue, hostility and sadness can be explained by the fact that all of them have a large number of moods consequently their feature space may be less sparse, spanning a variety of topical and linguistic contexts in Twitter posts. On the other hand, the worst performing classes, e.g., guilt, shyness and attentiveness, are also the ones with fewer corresponding moods. Hence it is possible that their feature spaces are rather sparse due to the limited contexts they are typically used in on Twitter. Moreover a qualitative study of the posts that belong to these classes tend to indicate significant degrees of sarcasm or irony in them e.g., for the guilt affect class: I hate when ppl read too deep into ur tweet and think it s about them... damn.. #guilty ; and for the attentiveness affect class: If a tomato is a fruit does that mean ketchup is a smoothie? #suspicious. Due to such contextual mismatch between content and the labeled affect, the classifier performs worse for these classes. Affect class Mean F1 Affect class Mean F1 Joviality Fear Fatigue Guilt Hostility Surprise Sadness Shyness Serenity Attentiveness Self-assurance Table 2. Mean F1 measures of 11 affect classes. What the classification results indicate in general is that, the manner in which the various affect classes are used on Twitter (via explicit mood hashtags) has a significant impact on the performance of the classifier. Moreover, we established earlier that different moods have different valence and activation measures (e.g., angry and frustrated are both negative moods, but angry indicates higher arousal than frustrated). These differences make the context of affect manifestation widely diverse in turn making affect classification in social media a challenging problem. UTILTY OF UNDERSTANDING EMOTION: CASE STUDY Finally, we discuss a direct application area where these mood and affect expression dynamics can be utilized to promote self-reflection and social awareness. Interest in developing applications to encourage and promote psychological well-being of individuals is not new to the HCI and health informatics communities [24, 25]. However there is exclusive value to utilizing expressed emotion on social media like Twitter in applications related to well-being. The value can be thought to be twofold. First, since Twitter allows us to track the emotions of individuals in a large time-scale (years), we can enable them observe their differences and changes in behavior over time, that could be indicative of a potential mental disorder (selfreflection). Second, since Twitter already has a social network associated with it, individuals are likely to obtain better social support from their contacts, than a different social platform, when it comes to helping themselves in tackling a health-related issue (social support and awareness). Motivated in this light, we present how mood expressions can be used to mine nuanced behavioral changes of individuals over time, in the light of major life events (e.g., birth of a child, loss of job, death related bereavement etc.). We propose a variety of measures to quantify such change over time, and reveal how emotion based measures can be used to forecast major events in people s lives that can, in turn, promote social and health-related well-being. Data Around A Major Life Event: Child-birth In this paper, we focus on a particular major event in a person's life: birth of a child. For the purpose, our population of interest comprises new mothers, i.e., female Twitter users who are likely to have given birth to a child in a given timeframe. Note that we focus on new mothers here, although we have observed that there is a fair percentage of Twitter posts from fathers right after their child-birth. The reason for this is that we believe new mothers experience a significant change in their lifestyle and habits in the postnatal period, compared to the fathers. Identifying new mothers on social media based on their posts, and in the absence of self-reported gender can be a challenging problem. Hence we follow an iterative approach involving first constructing a candidate set of likely new mothers based on filtering via several hardcrafted queries on the Twitter stream, gender inference, and finally identifying a set of high probability new mothers using ratings from Amazon's mechanical turkers. We discuss these steps as below: 1. We first construct a list of several hand-crafted queries to search the Twitter Firehose stream for candidate users likely to be posting about their child-birth. For the convenience of consistent prenatal and postnatal behavioral comparison of new mothers, we focus on searching the Firehose stream over a fixed two month period between May 1, 2011 and Jun 30, The different search queries included: after labor born, arrival baby boy/girl, birth pounds/inches, its a boy/girl born. It is worth mentioning here that these queries have been triggered by our observations that users tend to talk about the labor related to the final phase of their pregnancy, as well as tend to report on the physical details of their newborn child, including their gender: his/her weight and height. This phase yielded us a candidate set of 483 users. 2. Since we are interested in new mothers only, inferring the gender of the above constructed candidate user set was an important step. Twitter does not provide a facility for users to report on their gender; hence we had to rely on cues obtained from their self-declared name to infer the gender of the users in our candidate set. To this end, we utilized a simple lexicon based approach that attempts to find a match of the firstname of the Twitter user to a large dictionary of names collated from the United States

8 Census data, as well as a publicly available corpus of Facebook users names and self-reported gender. In a sense, because of the cross-cultural nature of Facebook, the dictionary amalgamating it and Census work fairly well in inferring gender of the Twitter users. We tested the accuracy of this inference mechanism by randomly selecting 100 users and labeling them manually the lexicon-driven gender inference mechanism yielded around 83% accuracy. In this manner, post gender identification, we obtained a smaller set of likely new mothers comprising 177 users. 3. In the final step, we use the gender-inferred set of likely new mothers in a mechanical turk setup, in order to rule out cases of false positives, and thereby obtain a high precision dataset. For the purpose, we showed each turker (min. 95% approval rating, English language proficient, and familiar with Twitter) a set of 10 Twitter posts from each user in our candidate set, such that 5 posts were posted right before the child-birth indicative post, and 5 after it. We hoped that this would give the turker some contextual cues to decide on whether the author of those posts appeared to be legitimately a new mother. Additionally, we also showed the Twitter profile bio, picture and a link to their actual Twitter profile for each user. The specific question involved responding to a yes/no/maybe multiple-choice type question per user, to evaluate if the user was a new mother. We thus collected five ratings per user from the turkers, and used the majority rating as the correct label (Fleiss-Kappa was 0.69). For our final dataset, we considered the users with the yes label and it consisted of 85 new mothers. Finally, for each of these 85 new mothers, we track their Twitter timelines in the Firehose stream to collect all of their posts in two 5-month periods, corresponding to prenatal and postnatal phases around child-birth (Dec 1, 2010 Apr 30, 2011 and Jul 1, 2011 Nov 30, 2011, respectively). In the following subsection, we will present some comparisons of the prenatal and postnatal behavioral trends along a number of emotional measures. Measures to Detect Behavioral Change We propose three different measures to quantify the behavioral change of the new mothers. These measures are motivated from the emotion studies discussed in the previous two sections. Note that our measures are crafted to determine behavioral differences that would reflect change in lifestyle in general, rather than aggregated volumes: hence they capture differences over the course of a day (24 hour periods). Volume. Our first measure called volume is based on the average normalized number of posts made per hour by the new mothers on a 24 hour period, over the prenatal and postnatal periods respectively. Positive Affect (PA). We define a measure of positive affect (PA) of the new mothers per hour during a 24 hour interval, during the prenatal and postnatal periods respectively. To compute PA, we utilize the psycholinguistic lexicon LIWC and focus on the words in the positive emotion category. Given a post from a new mother posted during a certain hour of the day, thereafter we perform a simple word spotting exercise to determine the fraction of words that match the words LIWC s positive emotion category. This fraction gives the measure of PA, when averaged over all posts for all mothers, per hour, corresponding to the prenatal and postnatal periods. Negative Affect (NA). Like PA, we also define a measure of negative affect (NA) averaged over all mothers per hour, in a typical 24 hour interval, during the prenatal and postnatal periods respectively. We again utilize LIWC for its negative affect categories: negative emotion, anger, anxiety, sadness. Based on the same word spotting technique, we measure NA per hour, averaged over all posts for all mothers corresponding to the prenatal and postnatal periods, over any typical 24 hour interval in a day. Prenatal and Postnatal Behavior Comparison Based on the three measures defined above, we now study the behavioral differences of new mothers, in the prenatal and postnatal periods. In particular, we are interested to Figure 6. Volume (average normalized post count), PA and NA for the set of 85 new mothers during the prenatal and postnatal periods. All measure demonstrate significant change, based on one-tail paired t-tests (p<0.0001).

9 observe how behavior changes during the postnatal period, with respect to the prenatal phase. Figure 6 shows the average trends of the measures in the prenatal and postnatal periods for all 85 new mothers. The results demonstrate that overall there is considerable change in behavior, as quantified by these measures (statistically significant based on one-tail paired t-tests). There is a general decrease in volume (normalized post counts) over a typical day in the postnatal period compared to prenatal. This probably reflects that mothers are experiencing busy times alongside the rather frequent maternity blues. Besides lower post volume on a social media site like Twitter might also indicate social withdrawal post childbirth. Additionally, there is high volume during the prenatal period in the early hours of the morning: a feature often typical of pregnant mothers suffering from morning sickness. This feature however disappears during the postnatal phase. We also observe a decrease in PA over the course of a day: likely because of a mother s physical, mental and emotional exhaustion [17], as well as sleep deprivation typical of parenting a newborn (notice the low PA in the early course of a day). Finally, in the case of NA, we observe that the postnatal period shows significantly high variance throughout the day, compared to that in the prenatal phase: an indicator of mood swings for the new mothers [7] as well as of increased anxiety or being overwhelmed frequently but inconsistently. We observe (not shown for interest of space) further significant behavioral trend change for new mothers for another set of measures that encompass their social interactions on Twitter (measures motivated from [4]): including number of re-tweets (RTs), number directed towards another user, number of links shared, as well as number of questions asked. We encounter similar finding here too: a general decrease in social interactions. To summarize, a major life event like child-birth triggers considerable behavioral change for new mothers in the postnatal period, compared to that during the prenatal phase; and social media activity can provide valuable emotional indicators to track these changes systematically. These findings also bear the potential to develop a prediction framework that can utilize the social media based emotional measures and forecast erratic behavior change in the near future for new mothers. DISCUSSION In this paper, we provided an elaborate exploration of the space of emotion expression in social media, in the light of how to measure them, how to infer them automatically when explicit signals are not available, and finally a utilitydriven case study to investigate how such measurement and inference of emotion can be put to practical use. Our broad findings extend existing psycholinguistics and psychology literature in several ways. Considerable efforts in prior research has focused on studying the circumplex model of moods; however studying the various extents of their usages by different individuals at a large-scale has still been a challenge due to the expensive mechanisms of collecting such data based on limited user studies. As Twitter continues to evolve as a mechanism of human expression, we have taken an effort to reveal dynamics of the circumplex of moods frequently used on Twitter, not only via their measures of valence and activation, but through their respective usages, linguistic diversity, diurnal patterns etc. as well. Our work in emotion exploration and measurement in social media provides ample scope for extensions and future research directions. A limitation of current work is that we have analyzed emotional behavior across all Twitter users in general, without making distinctions between different user types. For instance, one could conjecture that the mechanisms of mood expression will be different for elite users compared to ordinary individuals. Additionally there is likely to be differences in mood expression across cultures, demographics such as race, gender, social status, educational background and so on. Furthermore, studying the dynamics of the emotions as well as their utility in real world could reveal the susceptibility of various population segments to emergency situations an aspect that can benefit governmental agencies in particular. Finally, as a utility application domain for the insights obtained from measuring and inferring human moods in social media, we explored the topic of behavioral change of individuals around major life events. While there could be an entire diverse range of major life events that can trigger behavioral change, we have explored only one domain: postnatal behavior of new mothers. Though limited in its scope for the time-being, one could possibly expand this work to studying a variety of emotional disorders while investigating postnatal behavior change, such as postpartum depression (PPD), postnatal psychosis and so on. Plausibly, emotional markers obtained from social media and social networks data can enable better diagnosis of these disorders, at the same time provide mechanisms to new mothers to track their behavioral change for their health-related well-being. Opportunities also remain in terms of venturing out to other types of life events tracking whose behavioral change via social media activity can be fruitful. These include loss of a job or financial instability (to track population scale unemployment dissatisfaction, or economic indicators); death related grief and bereavement or loss of safety after a trauma (to help individuals cope with surprising emotions and improve their quality of life) and so on. In the coming years, from the perspective of HCI research, these could help us design emotion-aware interfaces: wherein a user's online social experience is tuned to her emotional state, emotional needs, requirement for social support as well as to act as a self-narrative and self-reflective feedback tool. CONCLUSIONS In this paper, we studied a variety of moods that frequent Twitter posts, from the point of view of their measurement, their inference in the absence of explicit signals, as well as

10 their potential in catering to practical real-world application scenarios that can promote social and personal wellness of individuals. We used the dimensions of valence and activation to represent moods in the circumplex model and studied the topology of this space with respect to mood usage, linguistic diversity, and activity. In this manner we provided naturalistic validation of the circumplex using social media data, and extended the conceptualization of human emotion at scale. Next, through our automatic affect classifier, we were able to expand emotion measurement and tracking to contexts where explicit mood signals might not be available: a contribution that can help recommender and search interfaces greatly. Finally, we investigated a case study in detecting behavioral changes of individuals (new mothers) around a major life event (child-birth), via their shared content on social media. Through this study we have laid the foundation of a line of HCI research that can utilize emotion signals from online activity to predict and forecast anomalous behavior or change at the personalized level of an individual, at the same time at a collective scale involving large populations. ACKNOWLEDGMENTS We thank Michael Gamon for help with the affect classification task, and Eric Horvitz for fruitful discussions. REFERENCES 1. Bollen, J., Mao, H., & Pepe, A. (2011). Modeling Public Mood and Emotion: Twitter Sentiment and Socio- Economic Phenomena. In Proc. ICWSM Bradley, M.M., & Lang, P.J. (1999). Affective norms for English words (ANEW). Gainesville, FL. The NIMH Center for the Study of Emotion and Attention. 3. Davidov, D., Tsur, O., & Rappoport, A. (2010). Enhanced Sentiment Learning Using Twitter Hashtags and Smileys. Proceedings of the 23rd International Conference on Computational Linguistics Posters, (August), De Choudhury, M., Counts, S., & Gamon, M. (2012). Not All Moods are Created Equal! Exploring Human Emotional States in Social Media. In Proc. ICWSM 2012, to appear. 5. De Choudhury, M., Gamon, M., and Counts, S. (2012). Happy, Nervous or Surprised? Classification of Human Affective States in Social Media. In Proc. ICWSM 2012, to appear. 6. Diakopoulos, N., & Shamma, D. (2010). Characterizing debate performance via aggregated twitter sentiment. In Proc. CHI Edhborg, M., Lundh, W., Seimyr, L., & Widstrom, A- M. (2001). The long-term impact of postnatal depressed mood on mothers + child interaction: a preliminary study. In Journal of Reproductive and Infant Psychology 19: Ekman, P. (1973). Cross-cultural studies of facial expressions. In P. Ekman (Ed.), Darwin and facial expression: A century of research in review (pp ). 9. Esuli, A., & Sebastiani, F. (2006). SentiWordNet: A publicly available lexical resource for opinion mining. Proceedings of LREC (Vol. 6, p ). 10. Golder, S. A., & Macy, M. W. (2011). Diurnal and Seasonal Mood Vary with Work, Sleep and Daylength Across Diverse Cultures. Science. 30 Sep Kleinginna, P.R., & Kleinginna, A.M. (1981). A catagorized list of motivation definitions with a suggestion for consensual definition. Motivation and Emotion, Kramer, A. D. I. (2010). An unobtrusive behavioral model of gross national happiness. In Proc. CHI Mehrabian, Albert (1980). Basic dimensions for a general psychological theory. pp Mihalcea, R., & Liu, H. (2006). A corpus-based approach to finding happiness. In Proceedings of computational approaches for analysis of weblogs, AAAI Spring Symposium. 15. Mishne, G. (2005). Experiments with mood classification in blog posts. In Style the 1st Workshop on Stylistic Analysis Of Text For Information Access, at SIGIR Mishne, G. & de Rijke, M. (2006). Capturing Global Mood Levels using Blog Posts. In AAAI 2006 Spring Symposium on Computational Approaches to Analyzing Weblogs. 17. O Hara, M.W. (1995). Postpartum Depression: Causes and Consequences. New York: Springer-Verlag. 18. Ortony, A., & Turner, T. J. (1990). What's basic about basic emotions? Psychological Review, 97, Pang, B., Lee, L., & Vaithyanathan, S. (2002). Thumbs up?: sentiment classification using machine learning techniques. In Proc. EMNLP 2002, Vol Russell, James A. (1980). A circumplex model of affect. J. of Personality and Social Psychology: 39, Tellegen, A. (1985). Structures of mood and personality and their relevance to assessing anxiety, with an emphasis on self-report. In A. H. Tuma & J. D. Maser (Eds.), Anxiety and the anxiety disorders (pp ). 22. Watson, D., & Clark, L. A. (1994). The PANAS-X: Manual for the positive and negative affect schedule- Expanded Form. Iowa City: University of Iowa. 23. Wiebe, J., Wilson, T., Bruce, R, Bell, M., & Martin, M. (2004). Learning subjective language. Computational Linguistics, 30 (3). 24. Anhoj, J., & Jensen, A-H. (2004). Using the Internet for life style changes in diet and physical activity: a feasibility study. In J Med Internet Res 8; 6(3):e Mamykina, L., Mynatt, E. et al. (2008). MAHI: investigation of social scaffolding for reflective thinking in diabetes management. In Proc. CHI 2008.