Author Profiling on Social Media: An Ensemble Learning Model using Various Features Notebook for PAN at CLEF 2019 Youngjun Joo and Inchon Hwang Department of Computer Engineering Yonsei University, Seoul, Korea {chrisjoo12, ich0103}@gmail.com Abstract We describe our participation in the PAN 2019 shared task on author profiling, determine whether a tweet’s author is a bot or a human, and in case of human, identify author’s gender for English and Spanish datasets. In this paper, we investigate the complementarities of both stylometry methods and content- based methods, putting forward various techniques for building flexible features. Acting as a complement to these methods, we investigate an ensemble learning method paves the way to improve the performance of AP tasks. Experimental re- sults demonstrate that the ensemble method by the combination of the stylometry methods and content-based methods can more accurately capture the author pro- files than traditional methods. Our proposed model obtained 0.9333 and 0.8352 of accuracy in the bot and gender identification tasks for English test dataset re- spectively. 1 Introduction The Author profiling (AP) deals with the classification of shared content in order to predict general or demographic attributes of authors such as gender, age, personality, native language, or political orientation, among others [17]. Being able to infer an au- thor’s profile has wide applicability and has proved to be advantageous in many areas such as marketing, forensics, and security, etc. Broadly speaking, the approaches that tackle AP view the task as a multi-class or single-label classification problem, when the set of the class label is known a priori [20]. Thus, AP is modeled as a classification task, in which automatic detection methods have to assign labels (e.g., male, female) to objects (texts). Consequently, most work has been devoted to determining a suitable set of features to deal with the task on the writing profile of authors. In the 2019 shared AP task on PAN dataset [2], the goal is to infer whether the author of a Twitter feed is a bot or a human and to profile the author’s gender in case of human [16]. Both training and test data is provided in two different languages: English, Spanish. 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. In order to predict bot and gender, we propose the complementarities of both sty- lometry methods and content-based methods, putting forward various techniques for building flexible features (basic count features, psycholinguistic features, TF-IDF, Doc2vec). Acting as a complement to these features, we also investigate an ensemble learning method combining classification methods based on various features and BERT model paves the way to improve the performance of AP tasks. 2 Related Works Approaches for predicting an AP can be broadly categorized into two types of methods: (1) stylometry methods which aim to capture an author’s writing style using different statistical features (e.g., functional words, POS, punctuation marks, and emoticons), (2) content-based methods that intend to identify an author’s profiles based on the content of the text (e.g., bag of words, words n-gram, term vectors, TF-IDF n-grams, slang words, emotional words), and topics discussed in the text (e.g., topic models such as LDA, PLSA ) [5]. According to the PAN1 competitions, most successful works for AP in social media have used combinations of these two kinds of features. Every author’s writing style can be used to identify an author’s attributes. In previ- ous studies, style based features were used to predict the author’s attributes, age, and gender [3,6,13,18,22]. In these methods, lexical word-based features represent text as a sequence of tokens forming sentences, paragraphs, and documents. A token can be the numeric number, alphabetic word or a punctuation mark. Plus, these tokens are used to statistics such as average sentence length, average word length, a total number of words and a total number of unique words, etc. Also, character-based features consider the text as a sequence of characters. Content-based methods employ specific words or special content which are used more frequently in that domain than in other domains [23]. These words can be chosen by correlating the meaning of words with the domain [8, 23] or selecting from corpus by frequency or by other feature selection methods [1]. An analysis of information gain presented in [19] showed that the most relevant features for gender identification are those related with content words (e.g., linux and office for identifying males, whereas love and shopping for identifying females). Recently, some works have used deep learning models and learning method of rep- resentations for AP [7, 9, 11, 21]. [7] used the approach based on subword character n-gram embeddings and deep averaging networks (DAN). [9] used the model consists of a bi-RNN implemented with a Gated Recurrent Unit (GRU) combined with an Atten- tion mechanism. [11] proposed two models for gender identification and the language variety identification of four languages that consist of multiple layers to classify an author’s profile trait with neural networks. 1 https://pan.webis.de/ Figure 1. Illustration of text preprocessing and postprocessing. 3 Proposed Approach 3.1 Text Preprocessing and Postprocessing The preprocessing of the text data is an essential step as it makes the raw text ready for applying machine learning algorithms to it. The objective of this step is to clean noise those are less relevant to detect the AP on the texts. We, at first, aggregate tweet posts published by an individual user into one docu- ment before training to alleviate the shortcomings of short texts. In order to utilize most of the information in text, we perform some transforming tasks of the short texts (i.e., XML parsing, contradictions unfolding, text tokenizing, stemming, lemmatization, and removing stopwords). Also, for utilization and word-level representation on most of the text information, we perform spell correction for informal words using SymSpell library1 , word segmentation for splitting hashtags using WordSegment library2 , and an- notation (surround or replace with special tags such as , , , , or ) as a text postprocessing task (see Figure 3). 3.2 Basic Count Features Previous works on AP tasks explore lexical, syntactic, and structural features. Lexi- cal features are used to measure the habit of using characters and words in the text. The commonly used features in this kind consist of the number of characters, word, a frequency of each type of characters, etc. Syntactic features include the use of punc- tuations, part-of-speech (POS) tags, and functional words. Structural features represent how the author organizes their documents or other special structures such as greetings or signatures. As shown Table 1, we construct a basic count feature set including punctuation char- acters (e.g., question marks, and exclamation marks) and other features (e.g., average syllable per word, functional word count, special character count, capital ratio, etc.). 1 https://github.com/wolfgarbe/SymSpell 2 https://github.com/grantjenks/python-wordsegment Table 1. List of basic count features extracted from the Twitter. Average sentence lenght by character Average sentence lenght by word Average syllable per word Average word length Capital ratio Number of functional words Number of special characters Character count Word count Hashtag count Direct tweet count Repeat punctuation count Punctuation count URL count Positive emoji count Negative emoji count Emoji count Slang words count Stopword count tag count tag count tag count