Sparse Coding in Authorship Attribution for Polish Tweets

Proceedings of Recent Advances in Natural Language Processing, pages 409–417, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_048 409 Sparse Coding in Authorship Attribution for Polish Tweets Piotr Grzybowski Wrocław University of Science and Technology Wrocław, Poland p.grzybowski2@gmail.com Ewa Juralewicz Wrocław University of Science and Technology Wrocław, Poland evajuralewicz@gmail.com Maciej Piasecki Wrocław University of Science and Technology Wrocław, Poland maciej.piasecki@pwr.edu.pl Abstract The study explores application of a simple Convolutional Neural Network for the problem of authorship attribu- tion of tweets written in Polish. In our solution we use two-step compres- sion of tweets using Byte Pair Encod- ing algorithm and vectorisation as an input to the distributional model gen- erated for the large corpus of Polish tweets by word2vec algorithm. Our method achieves results comparable to the state-of-the-art approaches for the similar task on English tweets and ex- presses a very good performance in the classification of Polish tweets. We tested the proposed method in relation to the number of authors and tweets per author. We also juxtaposed results for authors with different topic back- grounds against each other. 1 Introduction The problem of authorship attribution is one of the major areas of text classification. However, the issue is usually undertaken in the context of longer texts, such as book fragments, jour- nal articles or emails. In recent years, mass media channels started playing a huge role in social life and made it possible to participate in global discussions, especially through social media. Twitter is one of the most influential social media platforms where anyone can share their opinion in form of a short text. Along with the growing influence of such platforms and limited user verification possibilities, im- portance of verifying the authorship of tweets and other short texts published on social me- dia platforms has grown considerably. Such need is also motivated by moral responsibility of providing reliable media channels and dis- criminating propaganda messages. Inspired by the results obtained in (Shrestha et al., 2017) as well as the simplicity of their method, we decided to verify its abilities in terms of classifying tweets written by selected Polish influencers. We treat the problem as a multiclass classification of texts. Our method differs from the original approach in terms of a chosen method for text encoding and next the way of obtaining its distributional vector rep- resentation as well as the scope of usage of data prepared in the processing pipeline, i.e. we use different data, unrelated to that used for clas- sification, to train the distributional model. 2 Related Works A big part of research done in terms of au- thorship attribution is relevant for big chunks of texts, where the sample of author’s writ- ing is relatively big (Gollub et al., 2013), (Frantzeskou et al., 2007), (Koppel et al., 2011). Recently, a lot of work has been done regarding authorship attribution of short texts, especially tweets due to their accessibil- ity. The problem is challenging in compari- son to the classification of longer texts as it is harder to …

Content truncated to 500 words