A Deep Learning Ensemble Approach to Gender Identification of Tweet Authors

2017

Author profiling is a field within Natural Language Processing, in addition to being a

sub-field of the broader research area concerning authorship analysis. It aims to classify

personal traits of authors, such as gender and age, based on their writing style.

It is of growing importance with applications within fields such as forensics and marketing

for identifying characteristics of perpetrators and customers, respectively. The

emergence of social media platforms, such as Twitter, has resulted in a major increase

in textual user-generated content publicly available for linguistic studies. Additionally,

the informal language present in tweets provides linguistic material reflecting people s

everyday usage of language.

Though representation learning using deep learning has shown much promise, most of

the work within author profiling research in recent years has been based on the combination

of expensive manual feature engineering, representations such as Bag of Words,

and traditional machine learning methods exemplified by Support Vector Machines and

Logistic Regression. In this thesis we show that better gender-identifying feature representations

of English tweets can be learned using deep learning approaches.

We propose three classification systems, focusing on different granularities of text: a

character-level Convolutional Bidirectional Long Short-Term Memory (LSTM), a word-level

Bidirectional LSTM using Global Vectors (GloVe), and a more traditional document-level

system utilizing a feedforward network and Bag of Words of n-grams as first-level

representation. Furthermore, we propose using stacking to leverage the individual predictive

powers of the sub-models in a combined effort. The experiments reveal that the

word-level model outperforms the other sub-models, as well as the baseline models consisting

of Logistic Regression, Naïve Bayes and Random Forest. The best performance is

achieved by combining the character-level and word-level models, while the document-level

model dampens the combined performance.

NTNU