Predicting an author's age, gender and personality traits by analyzing his/her documents is important in forensics, marketing and resolving authorship disputes. Our system combines different styles, lexicons, topics, familial tokens and different categories of character n-grams as features to build a logistic regression model for four different languages: English, Spanish, Italian and Dutch. With this model, we obtained global ranking scores of 0.6623, 0.6547, 0.7411, 0.7662 for English, Spanish, Italian and Dutch languages respectively.
The PAN15 [
Researchers have approached this problem in a variety of ways on a number of different datasets. In order to capture the profile of Facebook users, Schwartz et al. used what they call an Open Vocabulary approach [7]. They used word n-grams and Latent Dirichlet Allocation (LDA) topics as features and compared their method with a Closed Vocabulary approach that used Linguistic Inquiry and Word Count (LIWC) word categories and found the Open Vocabulary method to be better for all of the personality traits as well as for age and gender.
Estival et al. [
In this paper, we built two separate models for age-gender and personality for each language. We used a wide range of features as described in Section 2 for profiling author’s age, gender and personality traits. We experimented with different combinations of these features with Logistic Regression as a classifier.
We started out by tokenizing author’s twitter data with the help of Ark Tweet NLP
tokenizer [
Twitter Style: Stylistic features capture the stereotypical style of a particular group of authors. Number of words, characters, question marks, exclamation marks, hashtags, user mentions (@), urls, all capitalized words, text and number combined tokens, average word length and average tweet length features were used to capture the writing style of authors.
Familial Tokens: Some of the familial tokens used by males and females are very distinct. Females are a lot more likely to use my hubby, my bf, my husband, my boyfriend, etc. than males. In the same way males will use words such as my wife, my girlfriend, my gf, etc. The presence of these tokens in authors’ tweets are strong indicative features to distinguish the gender of an author. We prepared a list of these tokens for all languages and used their normalized counts as features.
Categorical Char n-grams: Sapkota et al. [6] evaluated the predictive power of different subgroups of character n-grams in single and cross domain authorship attribution settings. They defined ten different character n-gram categories based on affixes, words and punctuation. Instead of using all the character n-grams, we defined similar categories and used those combinations of categories of char n-grams that gave us high accuracy. Categories like mid-word, beg-punct, multi-word, prefix, mid-punct and space-prefix proved to be good for this task.
LDA Topics: Many researchers have used LDA topics as features in order to predict gender, age group and personality of authors. We also used similarity of tweets with LDA topics as feature. For age and gender we clustered documents into eight topic groups and for personality we clustered into ten topic groups.
Age and Gender: We hypothesized that the same personality trait might have different patterns in authors from different age groups and gender. So, we used these as features in the determination of the personality of authors.
Apart from these features, we also experimented with the word categories in the
LIWC corpus. However, the addition of LIWC word lists degraded our system’s
performance and we dropped the LIWC word list from our final system. After obtaining these
features, we trained a multiclass logistic regression classifier with them. We used the
gensim [
Since, the training dataset is small, we performed our experiments through ten fold cross validation. We experimented with different combinations of the above defined features. Table 1 shows the combination of features that gave us the best results for age_gender and personality task. We created two separate models for age_gender and for personality classification. The personality model uses the age_gender model’s age_group and gender prediction as features.
Our system combined various style, lexicons, topics, familial tokens and categories of character n-grams features to build a final logistic regression classifier. Building a model with all of these features combined did not give good performance. But when combined selectively, these features boosted our system’s performance. In addition, we also found that the same type of features are not the ones that are strongly representative of authors’ profiles across different languages. For instance, the categories of character n-grams that were prominent across different languages were not the same. For English, Spanish and Italian, familial tokens did not improve the system performance, whereas for Dutch it is one of the key features. Similarly, Twitter specific style features are prominent across English, Spanish and Dutch but not in Italian. However, across all languages, word unigrams, bigrams and trigrams are important features, which illustrates that authors having similar attributes tend to use similar words. In conclusion, our system analyzes a wide range of features to profile author’s age-group, gender and personality traits and is reasonably successful in doing so.
We thank PAN15 organizers and committee members for organizing Author Profiling task. This research was funded by National Science Foundation CAREER grant 1462141. 6. Sapkota, U., Bethard, S., Montes, M., Solorio, T.: Not all character n-grams are created equal: A study in authorship attribution. In: Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 93–102. Association for Computational Linguistics, Denver, Colorado (May–June 2015), http://www.aclweb.org/anthology/N15-1010 7. Schwartz, H.A., Eichstaedt, J.C., Kern, M.L., Dziurzynski, L., Ramones, S.M., Agrawal, M., Shah, A., Kosinski, M., Stillwell, D., Seligman, M.E., et al.: Personality, gender, and age in the language of social media: The open-vocabulary approach. PloS one 8(9), e73791 (2013)