I’m a Belieber:
Social Roles via Self-identification and Conceptual Attributes

Charley Beller, Rebecca Knowles, Craig Harman
Shane Bergsma, Margaret Mitchell, Benjamin Van Durme
Human Language Technology Center of Excellence
Johns Hopkins University, Baltimore, MD USA
University of Saskatchewan, Saskatoon, Saskatchewan Canada
Microsoft Research, Redmond, Washington USA
charleybeller@jhu.edu, rknowles@jhu.edu, craig@craigharman.net,
shane.a.bergsma@gmail.com, memitc@microsoft.com, vandurme@cs.jhu.edu

Motivated by work predicting coarse-grained author categories in social media, such as gender or political preference, we explore whether Twitter contains information to support the prediction of fine-grained categories, or social roles. We find that the simple self-identification pattern “I am a  ” supports significantly richer classification than previously explored, successfully retrieving a variety of fine-grained roles. For a given role (e.g., writer), we can further identify characteristic attributes using a simple possessive construction (e.g., writer’s  ). Tweets that incorporate the attribute terms in first person possessives (my  ) are confirmed to be an indicator that the author holds the associated social role.

1 Introduction

With the rise of social media, researchers have sought to induce models for predicting latent author attributes such as gender, age, and political preferences [9, 19, 4, 22, 24]. Such models are clearly in line with the goals of both computational advertising [23] and the growing area of computational social science [6, 15, 16, 18, 14] where big data and computation supplement methods based on, e.g., direct human surveys. For example, Eisenstein et al. (2010) demonstrated a model that predicted where an author was located in order to analyze regional distinctions in communication. While some users explicitly share their GPS coordinates through their Twitter clients, having a larger collection of automatically identified users within a region was preferable even though the predictions for any given user were uncertain.

We show that media such as Twitter can support classification that is more fine-grained than gender or general location. Predicting social roles such as doctor, teacher, vegetarian, christian, may open the door to large-scale passive surveys of public discourse that dwarf what has been previously available to social scientists. For example, work on tracking the spread of flu infections across Twitter [12] might be enhanced with a factor based on aggregate predictions of author occupation.

We present two studies showing that first-person social content (tweets) contains intuitive signals for such fine-grained roles. We argue that non-trivial classifiers may be constructed based purely on leveraging simple linguistic patterns. These baselines suggest a wide range of author categories to be explored further in future work.

Study 1

In the first study, we seek to determine whether such a signal exists in self-identification: we rely on variants of a single pattern, “I am a  ”, to bootstrap data for training balanced-class binary classifiers using unigrams observed in tweet content. As compared to prior research that required actively polling users for ground truth in order to construct predictive models for demographic information [11], we demonstrate that some users specify such properties publicly through direct natural language.

Many of the resultant models show intuitive strongly-weighted features, such as a writer being likely to tweet about a story, or an athlete discussing a game. This demonstrates self-identification as a viable signal in building predictive models of social roles.

Study 2

In the second study we exploit a complementary signal based on characteristic conceptual attributes of a social role, or concept class [20, 1, 17]. We identify typical attributes of a given social role by collecting terms in the Google n-gram corpus that occur frequently in a possessive construction with that role. For example, with the role doctor we extract terms matching the simple pattern “doctor’s  ”.

2 Self-identification

Role Tweet
artist I’m an Artist….. the last of a dying breed
belieber @justinbieber I will support you in everything you do because I am a belieber please follow me I love you 30
vegetarian So glad I’m a vegetarian.
Table 1: Examples of self-identifying tweets.

All role-representative users were drawn from the free public 1% sample of the Twitter Firehose, over the period 2011-2013, from the subset that selected English as their native language (85,387,204 unique users). To identify users of a particular role, we performed a case-agnostic search of variants of a single pattern: I am a(n)  , and I’m a(n)  , where all single tokens filling the slot were taken as evidence of the author self-reporting for the given “role”. Example tweets can be seen in Table 1, examples of frequency per role in Table 2. This resulted in 63,858 unique roles identified, of which 44,260 appeared only once.11Future work should consider identifying multi-word role labels (e.g., Doctor Who fan, or dog walker).

# Role # Role # Role
29,924 little 5,694 man 564 champion
21,822 big 559 teacher
18,957 good 4,007 belieber 556 writer
13,069 huge 3,997 celebrity 556 awful
13,020 bit 3,737 virgin
12,816 fan 3,682 pretty 100 cashier
10,832 bad 100 bro
10,604 girl 2,915 woman
9,981 very 2,851 beast 10 linguist
Table 2: Number of self-identifying users per “role”. While rich in interesting labels, cases such as very highlight the purposeful simplicity of the current approach.

We manually selected a set of roles for further exploration, aiming for a diverse sample across: occupation (e.g., doctor, teacher), family (mother), disposition (pessimist), religion (christian), and “followers” (belieber, directioner).22Those that follow the music/life of the singer Justin Bieber and the band One Direction, respectively. We filtered users via language ID [2] to better ensure English content.33This removes users that selected English as their primary language, used a self-identification phrase, e.g. I am a belieber, but otherwise tended to communicate in non-English.

For each selected role, we randomly sampled up to 500 unique self-reporting users and then queried Twitter for up to 200 of their recent publicly posted tweets.44Roughly half of the classes had less than 500 self-reporting users in total, in those cases we used all matches. These tweets served as representative content for that role, with any tweet matching the self-reporting patterns filtered. Three sets of background populations were extracted based on randomly sampling users that self-reported English (post-filtered via LID).

Twitter users are empowered to at any time delete, rename or make private their accounts. Any given user taken to be representative based on a previously posted tweet may no longer be available to query on. As a hint of the sort of user studies one might explore given access to social role prediction, we see in Figure 1 a correlation between self-reported role and the chance of an account still being publicly visible, with roles such as belieber and directioner on the one hand, and doctor and teacher on the other.

Figure 1: Success rate for querying a user. Random.0,1,2 are background draws from the population, with the mean of those three samples drawn horizontally. Tails capture 95% confidence intervals.

The authors examined the self-identifying tweet of 20 random users per role. The accuracy of the self-identification pattern varied across roles and is attributable to various factors including quotes, e.g. @StarTrek Jim, I’m a DOCTOR not a download!. While these samples are small (and thus estimates of quality come with wide variance), it is noteworthy that a non-trivial number for each were judged as actually self-identifying.

Figure 2: Valid self-identifying tweets from sample of 20.

Indicative Language

Most work in user classification relies on featurizing language use, most simply through binary indicators recording whether a user did or did not use a particular word in a history of n tweets. To explore whether language provides signal for future work in fine-grain social role prediction, we constructed a set of experiments, one per role, where training and test sets were balanced between users from a random background sample and self-reported users. Baseline accuracy in these experiments was thus 50%.

Each training set had a target of 600 users (300 background, 300 self-identified); for those roles with less than 300 users self-identifying, all users were used, with an equal number background. We used the Jerboa [21] platform to convert data to binary feature vectors over a unigram vocabulary filtered such that the minimum frequency was 5 (across unique users). Training and testing was done with a log-linear model via LibLinear [8]. We used the positively annotated data to form test sets, balanced with data from the background set. Each test set had a theoretical maximum size of 40, but for several classes it was in the single digits (see Figure 2). Despite the varied noisiness of our simple pattern-bootstrapped training data, and the small size of our annotated test set, we see in Figure 3 that we are able to successfully achieve statistically significant predictions of social role for the majority of our selected examples.

Table 3 highlights examples of language indicative of role, as determined by the most positively weighted unigrams in the classification experiment. These results qualitatively suggest many roles under consideration may be teased out from a background population by focussing on language that follows expected use patterns. For example the use of the term game by athletes, studio by artists, mixtape by rappers, or jesus by Christians.

Figure 3: Accuracy in classifying social roles.
Role :: Feature (_Rank)
artist morning, summer, life, most, amp, studio
atheist fuck, fucking, shit, makes, dead, …, religion19
athlete lol, game, probably, life, into, …, team9
belieber justin, justinbeiber, believe, beliebers, bieber
cheerleader cheer, best, excited, hate, mom, …, prom16
christian lol, …, god12, pray13, …, bless17, …, jesus20
dancer dance, since, hey, never, been
directioner harry, d, follow, direction, never, liam, niall
doctor sweet, oh, or, life, nothing
engineer (, then, since, may, ), test9, -17, =18
freshman summer, homework, na, …, party19, school20
geek trying, oh, different, dead, been
grandma morning, baby, around, night, excited
hipster fucking, actually, thing, fuck, song
lawyer did, never, his, may, pretty, law, even, office
man man, away, ai, young, since
mother morning, take, fuck, fucking, trying
nurse lol, been, morning, …, night10, nursing11, shift13
optimist morning, enough, those, everything, never
poet feel, song, even, say, yo
rapper fuck, morning, lol, …, mixtape8, songs15
singer sing, song, music, lol, never
smoker fuck, shit, fucking, since, ass, smoke, weed20
solider ai, beautiful, lol, wan, trying
sophmore summer, >, …, school11, homework12
student anything, summer, morning, since, actually
teacher teacher, morning, teach, …, students7, …, school20
vegetarian actually, dead, summer, oh, morning
waitress man, try, goes, hate, fat
woman lol, into, woman, morning, never
writer write, story, sweet, very, working
Table 3: Most-positively weighted features per role, along with select features within the top 20. Surprising mother features come from ambigious self-identification, as seen in tweets such as: I’m a mother f!cking starrrrr.

3 Characteristic Attributes

Bergsma and Van Durme (2013) showed that the task of mining attributes for conceptual classes can relate straightforwardly to author attribute prediction. If one views a role, in their case gender, as two conceptual classes, male and female, then existing attribute extraction methods for third-person content (e.g., news articles) can be cheaply used to create a set of bootstrapping features for building classifiers over first-person content (e.g., tweets). For example, if we learn from news corpora that: a man may have a wife, then a tweet saying: …my wife… can be taken as potential evidence of membership in the male conceptual class.

In our second study, we test whether this idea extends to our wider set of fine-grained roles. For example, we aimed to discover that a doctor may have a patient, while a hairdresser may have a salon; these properties can be expressed in first-person content as possessives like my patient or my salon. We approached this task by selecting target roles from the first experiment and ranking characteristic attributes for each using pointwise mutual information (PMI) [5].

First, we counted all terms matching a target social role’s possessive pattern (e.g., doctor’s  ) in the web-scale n-gram corpus Google V2 [13]55In this corpus, follower-type roles like belieber and directioner are not at all prevalent. We therefore focused on occupational and habitual roles (e.g., doctor, smoker).. We ranked the collected terms by computing PMI between classes and attribute terms. Probabilities were estimated from counts of the class-attribute pairs along with counts matching the generic possessive patterns his   and her   which serve as general background categories. Following suggestions by Bergsma and Van Durme, we manually filtered the ranked list.66Evidence from cognitive work on memory-dependent tasks suggests that such relevance based filtering (recognition) involves less cognitive effort than generating relevant attributes (recall) see [10]. Indeed, this filtering step generally took less than a minute per class. We removed attributes that were either (a) not nominal, or (b) not indicative of the social role. This left fewer than 30 attribute terms per role, with many roles having fewer than 10.

We next performed a precision test to identify potentially useful attributes in these lists. We examined tweets with a first person possessive pattern for each attribute term from a small corpus of tweets collected over a single month in 2013, discarding those attribute terms with no positive matches. This precision test is useful regardless of how attribute lists are generated. The attribute term chart, for example, had high PMI with doctor; but a precision test on the phrase my chart yielded a single tweet which referred not to a medical chart but to a top ten list (prompting removal of this attribute). Using this smaller high-precision set of attribute terms, we collected tweets from the Twitter Firehose over the period 2011-2013.

4 Attribute-based Classification

Attribute terms are less indicative overall than self-ID, e.g., the phrase I’m a barber is a clearer signal than my scissors. We therefore include a role verification step in curating a collection of positively identified users. We use the crowdsourcing platform Mechanical Turk77https://www.mturk.com/mturk/ to judge whether the person tweeting held a given role Tweets were judged 5-way redundantly. Mechanical Turk judges (“Turkers”) were presented with a tweet and the prompt: Based on this tweet, would you think this person is a Barber/Hairdresser? along with four response options: Yes, Maybe, Hard to tell, and No.

We piloted this labeling task on 10 tweets per attribute term over a variety of classes. Each answer was associated with a score (Yes = 1, Maybe = .5, Hard to tell = No = 0) and aggregated across the five judges. We found in development that an aggregate score of 4.0 (out of 5.0) led to an acceptable agreement rate between the Turkers and the experimenters, when the tweets were randomly sampled and judged internally. We found that making conceptual class assignments based on a single tweet was often a subtle task. The results of this labeling study are shown in Figure 4, which gives the percent of tweets per attribute that were 4.0 or above. Attribute terms shown in red were manually discarded as being inaccurate (low on the y-axis) or non-prevalent (small shape).

From the remaining attribute terms, we identified users with tweets scoring 4.0 or better as positive examples of the associated roles. Tweets from those users were scraped via the Twitter API to construct corpora for each role. These were split intro train and test, balanced with data from the same background set used in the self-ID study.

Figure 4: Turker judged quality of attributes selected as candidate features for bootstrapping positive instances of the given social role.
Figure 5: Classifier accuracy on balanced set contrasting agreed upon Twitter users of a given role against users pulled at random from the 1% stream.

Test sets were usually of size 40 (20 positive, 20 background), with a few classes being sparse (the smallest had only 16 instances). Results are shown in Figure 5. Several classes in this balanced setup can be predicted with accuracies in the 70-90% range, supporting our claim that there is discriminating content for a variety of these social roles.

Conditional Classification

Figure 6: Results of positive vs negative by attribute term. Given that a user tweets …my lines … we are nearly 80% accurate in identifying whether or not the user is an actor.

How accurately we can predict membership in a given class when a Twitter user sends a tweet matching one of the targeted attributes? For example, if one sends a tweet saying my coach, then how likely is it that author is an athlete?

Using the same collection as the previous experiment, we trained classifiers conditioned on a given attribute term. Positive instances were taken to be those with a score of 4.0 or higher, with negative instances taken to be those with scores of 1.0 or lower (strong agreement by judges that the original tweet did not provide evidence of the given role). Classification results are shown in Figure 6.

5 Conclusion

We have shown that Twitter contains sufficiently robust signal to support more fine-grained author attribute prediction tasks than have previously been attempted. Our results are based on simple, intuitive search patterns with minimal additional filtering: this establishes the feasibility of the task, but leaves wide room for future work, both in the sophistication in methodology as well as the diversity of roles to be targeted. We exploited two complementary types of indicators: self-identification and self-possession of conceptual class (role) attributes. Those interested in identifying latent demographics can extend and improve these indicators in developing ways to identify groups of interest within the general population of Twitter users.


This material is partially based on research sponsored by the NSF under grants DGE-123285 and IIS-1249516 and by DARPA under agreement number FA8750-13-2-0017 (the DEFT program).


  • [1] A. Almuhareb and M. Poesio(2004) Attribute-based and value-based clustering: an evaluation. Cited by: 1.
  • [2] S. Bergsma, P. McNamee, M. Bagdouri, C. Fink and T. Wilson(2012) Language identification for creating language-specific twitter collections. Cited by: 2.
  • [3] S. Bergsma and B. Van Durme(2013) Using Conceptual Class Attributes to Characterize Social Media Users. Cited by: 3.
  • [4] J. D. Burger, J. Henderson, G. Kim and G. Zarrella(2011) Discriminating gender on twitter. Cited by: 1.
  • [5] K. W. Church and P. Hanks(1990) Word association norms, mutual information, and lexicography. Computational linguistics 16 (1), pp. 22–29. Cited by: 3.
  • [6] M. Conover, J. Ratkiewicz, M. Francisco, B. Gonçalves, F. Menczer and A. Flammini(2011) Political polarization on twitter.. Cited by: 1.
  • [7] J. Eisenstein, B. O’Connor, N. Smith and E. P. Xing(2010) A latent variable model of geographical lexical variation. Cited by: 1.
  • [8] R. Fan, K. Chang, C. Hsief, X. Wang and C. Lin(2008) LibLinear: a library for large linear classification. Journal of Machine Learning Research (9). Cited by: 2.
  • [9] N. Garera and D. Yarowsky(2009) Modeling latent biographic attributes in conversational genres. Cited by: 1.
  • [10] L. L. Jacoby, F. I. Craik and I. Begg(1979) Effects of decision difficulty on recognition and recall. Journal of Verbal Learning and Verbal Behavior 18 (5), pp. 585–600. Cited by: 3.
  • [11] M. Kosinski, D. Stillwell and T. Graepel(2013) Private traits and attributes are predictable from digital records of human behavior. Proceedings of the National Academy of Sciences. External Links: Document, Link, http://www.pnas.org/content/early/2013/03/06/1218772110.full.pdf+html Cited by: 1.
  • [12] A. Lamb, M. J. Paul and M. Dredze(2013) Separating fact from fear: tracking flu infections on twitter. Cited by: 1.
  • [13] D. Lin, K. Church, H. Ji, S. Sekine, D. Yarowsky, S. Bergsma, K. Patil, E. Pitler, R. Lathbury, V. Rao, K. Dalwani and S. Narsale(2010) New tools for web-scale n-grams. pp. 2221–2227. Cited by: 3.
  • [14] S. M. Mohammad, S. Kiritchenko and J. Martin(2013) Identifying purpose behind electoral tweets. WISDOM ’13, pp. 1–9. Cited by: 1.
  • [15] D. Nguyen, N. A. Smith and C. P. Rosé(2011) Author age prediction from text using linear regression. pp. 115–123. Cited by: 1.
  • [16] M. J. Paul and M. Dredze(2011) You are what you tweet: analyzing twitter for public health.. Cited by: 1.
  • [17] M. Paşca and B. Van Durme(2008) Weakly-Supervised Acquisition of Open-Domain Classes and Class Attributes from Web Documents and Query Logs. Cited by: 1.
  • [18] M. Pennacchiotti and A. Popescu(2011) Democrats, Republicans and Starbucks afficionados: user classification in Twitter. pp. 430–438. Cited by: 1.
  • [19] D. Rao, D. Yarowsky, A. Shreevats and M. Gupta(2010) Classifying latent user attributes in twitter. Cited by: 1.
  • [20] L. K. Schubert(2002) Can we derive general world knowledge from texts?. Cited by: 1.
  • [21] B. Van Durme(2012) Jerboa: a toolkit for randomized and streaming algorithms. Technical report Technical Report 7, Human Language Technology Center of Excellence, Johns Hopkins University. Cited by: 2.
  • [22] B. Van Durme(2012) Streaming analysis of discourse participants. Cited by: 1.
  • [23] J. Wortman(2008-05) Viral marketing and the diffusion of trends on social networks. Technical report Technical Report MS-CIS-08-19, University of Pennsylvania. Cited by: 1.
  • [24] F. A. Zamal, W. Liu and D. Ruths(2012) Homophily and latent attribute inference: inferring latent attributes of Twitter users from neighbors. Cited by: 1.