=Paper=
{{Paper
|id=Vol-2380/paper_154
|storemode=property
|title=Supervised Classification of Twitter Accounts Based on Textual Content of Tweets
|pdfUrl=https://ceur-ws.org/Vol-2380/paper_154.pdf
|volume=Vol-2380
|authors=Fredrik Johansson
|dblpUrl=https://dblp.org/rec/conf/clef/Johansson19
}}
==Supervised Classification of Twitter Accounts Based on Textual Content of Tweets==
https://ceur-ws.org/Vol-2380/paper_154.pdf
Supervised Classification of Twitter Accounts Based on
Textual Content of Tweets
Notebook for PAN at CLEF 2019
Fredrik Johansson
Swedish Defence Research Agency (FOI)
fredrik.johansson@foi.se
Abstract In our implemented system submitted to the bots and gender profil-
ing task of PAN 2019, we use a two-step binary classification approach in which
we classify accounts as being bot or not based on a combination of term oc-
currences and aggregated statistics fed to a random forest classifier. Accounts
classified as human are further distinguished as male or female through a logistic
regression classifier taking data-driven function words as input. We obtain highly
competetive bot and gender classification accuracies on English (0.96 and 0.84,
resepectively) while performing worse on Spanish (0.88 and 0.73, respectively).
1 Introduction
In the modern society a large number of communication and interactions take place on
various social media platforms such as Twitter and Facebook. Social media analytics is
being used for everything from finding out what people think about a specific brand or
product [16] to estimating citizens’ reactions to crisis events [2]. Profiling of authors (or
their corresponding user accounts) is an important part of the analytical process, since
the presence of e.g., Twitter bots otherwise can give a skewed view of things like the true
public opinion on a product or a political party. Likewise, it can be benificial to be able
to estimate whether a social media account is likely to belong to a man or a woman,
since gender is often an important variable for product marketers, political campaign
staff, or crisis managers to keep track of. In the PAN 2019 bots and gender profiling
task [10], the challenge is to given a Twitter feed, determine whether its author is a bot
or a human, and in case of a human, to identify the gender (male or female) of the author.
Since social media is highly multilingual, there is both an English and a Spanish subset
of the task. To further increase the complexity of the task, only the textual content of the
tweets is available, i.e., no metadata or other social network information are available
during neither training nor testing.
Our implemented solution is intended to be as generic as possible so that it is useful
also for other languages than English and Spanish, as well as being useful for other
types of Twitter accounts than the specific distribution of accounts present in this chal-
lenge. For this reason we have attempted to only include features which are likely to
Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons Li-
cense Attribution 4.0 International (CC BY 4.0). CLEF 2019, 9-12 September 2019, Lugano,
Switzerland.
generalize, rather than to only aim for as high predictive performance as possible on
the PAN dataset. The rest of this paper is structured as follows. We start by giving a
brief overview of related work in Section 2. In Section 3 we present the overall design
choices of the implemented solution, such as how we have designed the classification
problem to solve. Next, we are in Section 4 presenting the features and classifiers used
in our implemented system. The obtained experimental results are described in Section
5 and we present overall conclusions in Section 6.
2 Related Work
Twitter is a social media platform on which bots have been present for a long time.
There are several types of bots, ranging from spam bots to bots spreading malware and
automated accounts used for various kinds of information operations in political and
military conflicts [13]. An exhaustive overview of previous work on detecting bots on
Twitter is outside the scope of this paper, but most of them rely on machine learning
classifiers applied to metadata-based features such as account age, the used client, the
ratio between followers and friends, the ratio between tweets and retweets, and average
number of tweets per hour or day. One early example of such an approach in which the
authors classify a Twitter account as being either human, bot, or cyborg is the work by
Chu et al. [4]. In 2015, DARPA conducted a Twitter Bot Detection Challenge, in which
participating teams were supposed to develop and test the effectiveness of their bot de-
tection methods [12]. The participating teams used metadata-based features as those
already discussed, but also network-based features such as degree centrality, number of
accounts classified as bots in the cluster a user account belongs to, as well as content-
based features [12]. A last example of previous work on bot classification is [14]. They
use a total of 1150 features, including features related to user metadata, friends, net-
work, and timing. They also make use of content-based features such as frequency of
POS tags and features related to sentiment. However, they restrict their work to English-
speaking Twitter users.
Some previous work has also been devoted to gender classification of Twitter users.
One well-known example is [3], in which the authors use the textual content as well as
user metadata consisting of full name, user name, and the description from the user pro-
file. The authors make use of word- and character-level n-grams. There have also been
previous work on classifying Twitter accounts as male or female on PAN. However, this
challenge has involved posted tweets as well as user images.
3 Overall Design Choices
A first design choice to make is whether to track the classification problem as a multi-
class problem with the three possible classes male, female, and bot, or to treat it as
a two-stage binary classification problem with classes human, bot and male, female,
respectively. We have experimented with both versions, but have in the end decided to
model it as a two-stage binary classification problem. The main reason for this is that
we in our applied research work have several practical scenarios related to detection
of information operations in which we have a need for classifying Twitter accounts as
being bot or not, and a few security-related applications in which we need to separate
between male and female owners of Twitter accounts. However, we see few real-world
examples in which we would like to be able to do both kinds of classification in the same
scenario. Moreover, the task has been presented as a two-stage classification problem
in the task description at the PAN web page. This being said, it is not obvious that this
problem formulation will yield better results than a multi-class definition of the problem
would result in.
In many classification problems the instances are classified independently of each
other, but here the individual instances correspond to single tweets while the task is to
classify the accounts or authors creating these tweets. Hence, there are two options:
1. To classify tweet by tweet and to somehow weigh these individual classifications
into an aggregate classification of the Twitter account as such (e.g., by using ma-
jority voting).
2. To first calculate aggregate statistics from all tweets by a specific account and to
use these as features on which classifiers can be applied directly.
We have in this work used the latter approach as we think there are many signs of
Twitter bots which are not shown on the level of individual tweets, but which become
more clear on an aggregate level. An example of this is how similar all tweets from a
single Twitter account are. If all tweets are highly similar, this may be a good indicator
of this account being a bot, but tweet similarity is not an available feature if classifying
the tweets independently of each other. The specific features which have been used in
our implementation are described more thoroughly in Section 4.
4 Implemented Features and Classifiers
Our implemented features and classifiers used for distinguishing between bots and hu-
mans are first described in Section 4.1. Next, we describe the features and classifiers
used for distinguishing between male and female authors in Section 4.2.
4.1 Bot Classification
When designing our text-based bot vs. human features, we have thought about which
metadata we previously have found useful for classifying Twitter accounts as bot or
human in [13] and reasoned about how such features can be extracted or derived from
the textual content alone when no metadata is available. Some of the first identified
features on tweet level are tweet length (number of characters) and number of capital
letters, as these are thought to be potential indicators for e.g., very basic spam bots. We
then calculate aggregate statistics for max, min, avg, and std for an account based
on the counts from the tweet level, since the classification is supposed to take place on
account level rather than individual tweet level, as discussed previously. In the same
manner, we calculate statistics based on serveral other counts:
– Number of URLs (number of words starting with "http")
– Number of mentions (number of words starting with "@")
– Fraction of retweets (number of tokens starting "RT")
– Number of lowercase letters
For all these features we can treat the tweets from a specific account independently
from each other. However, we also think that bot behavior can be captured by mea-
suring the similarity among subsequent tweets. For this reason we use the ordering
in which the tweets from the same account appear in the dataset (assuming this cor-
responds to some time ordering) and calculate the edit distance (or more specifically
the Damerau-Levenshtein distance [5]) between all consecutive tweets. These distanes
are then aggregated into max, min, avg, and std features in the same way as the less
complex features.
Finally, an important clue to whether a Twitter account is automated or operated by
a human author is the actual wordings of the tweets. For example, tweets containing the
word "VIAGRA" or "SALE" are probably more likely to have been created by a spam
bot than a human. However, there is a great risk of overfitting to a particular type of
dataset (reducing the possibility for real-world usage) unless care is taken to regularize
or in other ways constrain the classifier’s reliance on specific words. We have chosen
to simply concatenate all tweets belonging to the same Twitter account and then apply
tf-idf to the most common unigrams and bigrams. We have experimented with various
choices of the number of features to extract from the tf-idf, but settled for 800, which
is quite restrictive. Moreover, we have found term presence to work at least equally
good as term frequencies, which further reduce the complexity of the extracted features.
A higher number of extracted features could potentially have given better results on
the unseen test set, but we decided to set it low in order to mainly keep words likely
to generalize well also to other datasets. More modern natural language processing
techniques such as word embeddings [7] or more sophisticated architectures such as
Long Short-Term Memories (LSTMs) [6] or Transformers [15] would likely have given
better results than a simple tf-idf, however, this have not been experimented with due
to the rather long time it takes to run even simple classifiers on the TIRA [9] virtual
matchines (on which the evaluation of the submitted solutions have to take place).
We have now presented all the used features for bot vs. human classification, but
there are also various choices to be made on how to combine the features into actual
classifications. One obvious way would be to combine all discussed features into a
single feature vector onto which a single classifier is applied. The feature vectors from
the tf-idf are, however, quite sparse compared to the aggregated statistical features and
we have found that we get better performance if we first train a separate classifier on the
tf-idf features only and then add the output from this first classifier as an extra feature
to the rest of the statistical features. We have experimented with both the crisp class
predictions as well as probability distributions over the different class labels and found
that using the probability distributions as an extra feature to work slightly better on
our hold-out test data. We have also experimented with various classifiers. In our final
implementation of the submitted system we have used a logistic regression classifier
(with regularization set to C = 1.0) for the tf-idf features and a random forest classifier
(with 500 estimators and a minimum of one sample per leaf) which have been applied
to the combination of the class probability output from the logistic regression classifier
and the aggregated statistical features.
4.2 Gender Classification
For the Twitter accounts that are classified as human in the first step, we are in the next
step applying a gender classifier intended to distinguish between male and f emale
authors. For this problem we have experimented with the same features as described
above. However, in the end we used only tf-idf, with the 300 most common lower-
cased word unigrams in the training set, as this performed better than if these features
were combined with any of the other features describe above. Upon inspection it can be
seen that these unigrams are basically corresponding to (data-driven) function words in
the English and Spanish languages, respectively. Function words have previously been
found to work well for gender classification, see e.g., [11]. Other features that have been
shown to work well for gender classification of e.g., blogs are part-of-speech (POS) tags
[1]. We have experimented with POS tags obtained using SpaCy1 , however, the POS
tagging takes some extra time and did not seem to provide any extra predictive power,
so in the end we only used the tf-idf features described above for the final submission.
These features are used as input to a logistic regression classifier (with regularization
set to C = 1.0).
5 Experimental Results
All the features and classifiers described in the last section have been implemented us-
ing scikit-learn [8]. In order to get an idea of what kind of features and classifiers that
work well for the PAN 2019 bots and gender profiling task, we initially set aside a ran-
domly selected subset of 10 percent of the English and Spanish Twitter accounts in the
available dataset, respectively. This gave very high initial accuracies and F1-scores also
for very basic algorithms. However, when discovering and utilizing the train/dev split
recommended by the PAN organizers on their web page, more reasonable results were
obtained. The reason for this difference is unknown to the author, but we assume it is
a consequence of how the datasets have been collected or constructed. The numbers
reported in this section have all been computed on the dev part of the dataset recom-
mended by the PAN organizers, unless it is explicitly mentioned that the numbers have
been achieved on the final submission through TIRA [9], i.e., pan19-author-profiling-
test-dataset1-2019-03-20 (here after referred to as testset1) and pan19-author-profiling-
test-dataset2-2019-04-29 (here after referred to as testset2).
When training the presented system on the train part and evaluating on the dev part
of the train/dev dataset obtained as part of the PAN task, the results summarized in
Table 1 were obtained.
When submitting the best performing version of the system to TIRA, the accuracies
presented in Table 2 were obtained for the different languages and the different classi-
fication problems. Based on these results we can say that the classification of accounts
as being bot or human is, as expected, an easier task than to discriminate between male
and female authors. Moreover, it is also shown that the implemented system performs
1
https://github.com/explosion/spaCy
Table 1. Accuracies obtained on the bots and gender profiling tasks on the local evaluation.
Classification
Language Tf-idf features Tf-idf + statistical features
task
Bots profiling en 0.914 0.948
Bots profiling es 0.865 0.892
Gender profiling en 0.752 0.713
Gender profiling es 0.648 0.617
Table 2. Accuracies obtained on the bots and gender profiling tasks on the TIRA test datasets.
Classification
Language testset1 testset2
task
Bots profiling en 0.958 0.960
Bots profiling es 0.806 0.882
Gender profiling en 0.792 0.838
Gender profiling es 0.606 0.728
significantly better on English than on Spanish for both the bots and gender predic-
tion tasks, despite that we have not used any resources or features that are specifically
designed for English.
If we compare the obtained accuracies to the other participating teams, it seems like
the accuracy of our submitted bot classifier is performing better than all other teams
on English, while it is only among the 15 best performing submissions on Spanish. On
gender profiling it is among the top-5 submissions on English, while it is only among
the 20 best performing submissions on Spanish. This clearly shows that our submit-
ted system works well on English, while it underperforms on Spanish (and have plenty
of room for improvement). When the obtained accuracy scores per language are aver-
aged we obtain an average score of 0.852, yielding a final ranking of place 9 out of 55
participating teams.
6 Conclusions
Our submission to the PAN 2019 bots and gender profiling task relies on a two-step
binary classification in which we first predict whether Twitter accounts are bots or not
(i.e., human), and next predict whether the owners of the accounts classified as human
are male or female. For both classifications we have made use of standard tf-idf features
and mildly regularized logistic regression classifiers, but for the bot or not classification
we have combined the output from the logistic regression classifier with many other
statistical features derived from the textual content of the tweets, including Damerau-
Levenshtein distance-based metrics and the number of retweets. These combined fea-
tures have been input to a random forest classifier, which produced the best performance
in our conducted experiments. In our performed local experiments we obtained accura-
cies of approximately 0.95 (English) and 0.89 (Spanish) on the bot classification task,
and approximately 0.75 (English) and 0.65 (Spanish) on the gender classification task.
When submitting the same models to TIRA and evaluating them on previously unseen
test data we received results that were inline with these numbers as well.
Since we have used quite constrained features and have based them on reasoning
about what kind of features that are likely to be useful for bot and gender classification,
we expect the implemented system to be useful also in real-world settings and not only
within the PAN competition. The basic nature of the implemented features and classi-
fiers leave plently of room for improvement by using more state-of-the-art embeddings
or deep learning architectures.
Acknowledgments
This work was supported by the R&D programme of the Swedish Armed Forces.
References
1. Argamon, S., Koppel, M., Pennebaker, J.W., Schler, J.: Mining the blogosphere: Age,
gender and the varieties of self-expression. First Monday 12(9) (2007)
2. Brynielsson, J., Johansson, F., Jonsson, C., Westling, A.: Emotion classification of social
media posts for estimating people’s reactions to communicated alert messages during
crises. Security Informatics 3(1), 7 (Aug 2014), https://doi.org/10.1186/s13388-014-0007-3
3. Burger, J.D., Henderson, J., Kim, G., Zarrella, G.: Discriminating gender on twitter. In:
Proceedings of the Conference on Empirical Methods in Natural Language Processing. pp.
1301–1309. EMNLP ’11, Association for Computational Linguistics, Stroudsburg, PA,
USA (2011), http://dl.acm.org/citation.cfm?id=2145432.2145568
4. Chu, Z., Gianvecchio, S., Wang, H., Jajodia, S.: Detecting automation of twitter accounts:
Are you a human, bot, or cyborg? IEEE Trans. Dependable Secur. Comput. 9(6), 811–824
(Nov 2012), http://dx.doi.org/10.1109/TDSC.2012.75
5. Damerau, F.J.: A technique for computer detection and correction of spelling errors.
Commun. ACM 7(3), 171–176 (Mar 1964), http://doi.acm.org/10.1145/363958.363994
6. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8),
1735–1780 (Nov 1997), http://dx.doi.org/10.1162/neco.1997.9.8.1735
7. Mikolov, T., Sutskever, I., Chen, K., Corrado, G., Dean, J.: Distributed representations of
words and phrases and their compositionality. In: Proceedings of the 26th International
Conference on Neural Information Processing Systems - Volume 2. pp. 3111–3119.
NIPS’13, Curran Associates Inc., USA (2013),
http://dl.acm.org/citation.cfm?id=2999792.2999959
8. Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M.,
Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D.,
Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine Learning in Python . Journal
of Machine Learning Research 12, 2825–2830 (2011)
9. Potthast, M., Gollub, T., Wiegmann, M., Stein, B.: TIRA Integrated Research Architecture.
In: Ferro, N., Peters, C. (eds.) Information Retrieval Evaluation in a Changing World -
Lessons Learned from 20 Years of CLEF. Springer (2019)
10. Rangel, F., Rosso, P.: Overview of the 7th author profiling task at pan 2019: Bots and gender
profiling. In: Cappellato, L., Ferro, N., Müller, H., Losada, D. (eds.) CLEF 2019 Labs and
Workshops, Notebook Papers. CEUR Workshop Proceedings. CEUR-WS.org (2019)
11. Schler, J., Koppel, M., Argamon, S., Pennebaker, J.: Effects of age and gender on blogging.
In: Computational Approaches to Analyzing Weblogs - Papers from the AAAI Spring
Symposium, Technical Report. vol. SS-06-03, pp. 191–197 (8 2006)
12. Subrahmanian, V.S., Azaria, A., Durst, S., Kagan, V., Galstyan, A., Lerman, K., Zhu, L.,
Ferrara, E., Flammini, A., Menczer, F.: The darpa twitter bot challenge. Computer 49(6),
38–46 (Jun 2016), https://doi.org/10.1109/MC.2016.183
13. Teljstedt, C., Rosell, M., Johansson, F.: A semi-automatic approach for labeling large
amounts of automated and non-automated social media user accounts. In: 2015 Second
European Network Intelligence Conference. pp. 155–159 (2015)
14. Varol, O., Ferrara, E., Davis, C.A., Menczer, F., Flammini, A.: Online human-bot
interactions:detection, estimation, and characterization. In: Proceedings of the Eleventh
International AAAI Conference on Web and Social Media (ICWSM 2017) (2017)
15. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L.,
Polosukhin, I.: Attention is all you need (2017), https://arxiv.org/pdf/1706.03762.pdf
16. Zhang, L., Liu, B.: Sentiment Analysis and Opinion Mining, pp. 1152–1161. Springer US,
Boston, MA (2017)