Author profiling is the analysis of a given set of texts in an attempt to uncover various characteristics of the author based on stylistic- and content-based features, or to identify the author. Characteristics analysed commonly include age and gender, though more recent studies have looked at other characteristics, like personality traits and occupation.Wiegmann, M., Stein, B. & Potthast, M. (2019). "Overview of the Celebrity Profiling Task at PAN 2019." CLEF.
Author profiling is one of the three major fields in automatic authorship identification (AAI), the other two being authorship attribution and authorship identification. The process of AAI emerged at the end of the 19th century. Thomas Corwin Mendenhall, an American autodidact physicist and meteorologist, was the first to apply this process to the works of Francis Bacon, William Shakespeare, and Christopher Marlowe. From these three historic figures, Mendenhall sought to uncover their quantitative stylistic differences by inspecting word lengths.Mikros, G.K., & Perifanos, K. (2013). "Authorship attribution in Greek tweets using author's multilevel n-gram profiles." 2013 AAAI Spring Symposium Series.
Although much progress has been made in the 21st century, the task of author profiling remains an unsolved problem due to its difficulty.
Techniques
Through the analysis of texts, various author profiling techniques can be applied to predict information about the author. For example, function words, as well as part-of-speech analysis, can be referenced to determine the author's gender and truth of a text.Koppel, M., Argamon, S., & Shimoni, A.R. (2013). "Automatically categorizing written texts by author gender." Literary and Linguistic Computing, 17, pg 401–412.
The process of author profiling usually involves the following steps:López-Monroy, A. P., Montes-y-Gómez, M., Escalante, H. J., Villaseñor-Pineda, L. & Stamatatos, E. (2015). "Discriminative subprofile-specific representations for author profiling in social media." In: Knowledge-Based Systems, 89, 134 – 147. Identifying specific features to be extracted from the text Building an adopted, standard representation (e.g. Bag-of-words model) for the target profile Building a classification model using a standard classifier (e.g. Support Vector Machines) for the target profile
Machine learning algorithms for author profiling have become increasingly complex over time. Algorithms used in author profiling include: Support Vector Machines Lundeqvist, E. & Svensson, M. (2017). "Author profiling: A machine learning approach towards detecting gender, age and native language of users in social media." In: Department of Information Technology. Naive Bayes classifiers Deep averaging networks,Franco-Salvador, M., Plotnikova, N., Pawar, N., & Benajiba, Y. (2017). "Subword-based deep averaging networks for author profiling in social media." CLEF. many layers in a cycle of machine learning that uses the mean of word embeddings within a textKurita, K. (2018). "Paper dissected: Deep unordered composition rivals syntactic methods for text classification explained." Machine Learning Explained. Long Short-Term memoryBsi, B. & Zrigui, M. (2018). "Deep learning techniques for author profiling in social media content." In: 31st IBIMA Conference.
In the past, author profiling was limited to physical documents, often in the form of books and newspaper articles. Different combinations of textual attributes belonging to the authors were identified and analyzed using author profiling, including lexical and syntactical features. Pioneering research in author profiling focused mostly on a single genre until the shift towards author profiling on social media and the Internet.Bilan, I. & Zhekova, D. (2016). "CAPS: A cross-genre author profiling system." CLEF. While attributes, such as content words and POS tags, are effective in author profile predictions on physical documents, their effectiveness in author profile predictions on digital texts is subjective and dependent on the type of online content being analyzed.
With the advances in technology, author profiling on the Internet has become increasingly common. Digital texts, such as social media posts, blog posts and emails, are now being used. This has sparked greater research efforts because of the advantages analysing digital texts can bring to sectors like marketing and business. Author profiling on digital texts has also enabled predictions of a wider range of author characteristics such as personality, income and occupation.Schler, J., Koppel, M., Argamon, S., & Pennebaker, J.W. (2005). "Effects of Age and Gender on Blogging." AAAI Spring Symposium: Computational Approaches to Analyzing Weblogs.
The most effective attributes for author profiling on digital texts involve a combinations of stylistic and content features. Author profiling on digital texts focuses on cross-genre author profiling, whereby one genre is used for training data and another genre is used for testing data, though both need to be relatively similar for good results.
There are some problems when performing author profiling techniques on online texts. These problems include: Wide variation in lengths of texts used Class imbalance in data
Author profiling and the Internet
The rise of the internet in the 20th to 21st century catalysed an increase in author profiling research, since data could be mined from the web, including social media platforms, emails and blogs. Content from the web have been analysed in tasks of author profiling to identify the age, gender, geographic origins, nationality and psychometric traits of web users. The information obtained has been used to serve various applications, including marketing and forensics.
Social media
The increased integration of social media in people's daily lives have made them a rich source of textual data for author profiling. This is mainly because users frequently upload and share content for various purposes including self-expression, socialisation, and personal businesses. The Social bot is also a frequent feature of social media platforms, especially Twitter, generating content that may be analysed for author profiling.Rangel, F., & Russo, P. (2019). "Overview of the 7th author profiling task at PAN 2019: Bots and gender profiling in Twitter." CLEF. While different platforms contain similar data they may also contain different features depending on the format and structure of the particular platform.
There are still limitations in using social media as data sources for author profiling, because data obtained may not always be reliable or accurate. Users sometimes provide false information about themselves or withhold information.Rosso, P., Rangel, F., Farías, I. H., Cagnina, L., Zaghouani, W., & Charfi, A. (2018). "A survey on author profiling, deception, and irony detection for the Arabic language." Language and Linguistics Compass, 12(4). As a result, the training of algorithms for author profiling may be impeded by data that is less accurate. Another limitation is the irregularity of text in social media. Features of irregularity include deviation from normal linguistic standards such as spelling errors, unstandardised transliteration as with the substitution of letters with numbers, shorthands, user-created abbreviations for phrases and et cetera, which may pose a challenge to author profiling.Gómez-Adorno, H., Markov, I., Sidorov, G., Posadas-Durán, J.-P., Sanchez-Perez, M. A., & Chanona-Hernandez, L. (2016). "Improving Feature Representation Based on a Neural Network for Author Profiling in Social Media Texts". In: Computational Intelligence and Neuroscience, pg 1–13. Researchers have adopted methods to overcome these limitations in training their algorithms for author profiling.
Facebook is useful for author profiling studies as a social networking service. This is because of how a social network may be built, expanded, and used for social action in the site.Dam, J. W. V., & Velden, M. V. D. (2015). "Online profiling and clustering of Facebook users". In: Decision Support Systems, 70, 60–72. In such processes, users share personal content that may be used for author profiling studies. Textual data is obtained from Facebook for author profiling from user's personal posts such as 'status updates'.Hsieh, F.C., Sandroni, R.F., & Paraboni, I. (2018). "Author Profiling from Facebook Corpora". LREC. These are acquired to produce a corpus in the selected language(s) for author profiling, to create either a bilingual or multilingual database of content words,Fatima, M., Hasan, K., Anwar, S., & Nawab, R. M. A. (2017). "Multilingual author profiling on Facebook". In: Information Processing & Management, 53(4), 886–904. which may then be used for author profiling.
In the context of Facebook, author profiling mainly involves English textual data, but also uses non-english languages that include: Roman Urdu, Arabic, Brazilian Portuguese, Spanish. While author profiling studies on Facebook have been predominantly for gender and age-group identification, there have been attempts to derive attributes to predict religiosity, the IT background of users, and even basic emotions (as defined by Paul Ekman) among others.Rangel, F., & Rosso, P. (2013). "Use of Language and Author Profiling: Identification of Gender and Age."
Sina Weibo is one of the few Asian social media platforms that contain texts in Asian languages to have been analysed for author profiling. Primary content of focus for author profiling on Weibo content include classical Chinese characters, hashtags, emoticons, kaomoji, homogenous punctuation, Latin sequences (due to the multilingualism of text) and even poetic formats. Particularly popular Chinese expressions, POS tags and word types are also tracked for author profiling.Zhang, W., Caines, A., Alikaniotis, D., & Buttery, P. (2015). "Predicting author age from Weibo microblog posts." LREC.
Author profiling for Weibo content requires algorithms different from those used for other social media platforms, mainly due to the linguistic differences between Mandarin Chinese and Western languages. For example, Chinese emotions involve Chinese characters describing the gesture or facial expression in brackets, such as: e.g. [哈哈] 'laughter', [泪] 'tears', [偷笑] 'giggle', [爱你] 'love', [心] 'heart'. This differs from the use of punctuation symbols for emoticons in Western languages, or the common use of the Unicode emojis in other platforms such as Facebook, Instagram, et cetera. Further, while there are around 161 western emoticons, there are around 2900 emoticons regularly used in mainland China for web content as in Weibo.Chen, L., Qian, T., Wang, F., You, Z., Peng, Q., & Zhong, M. (2015). "Age Detection for Chinese Users in Weibo." WAIM 2015, LNCS 9098, 83–95. To tackle these differences, author profiling algorithms have been trained on Chinese emoticons and linguistic features. For example, author profiling algorithms have been designed to detect Chinese stylistic expressions expressing formality and sentiment, in place of algorithms detecting English linguistic features such as capital letters.
As compared to other more popular, globalised platforms, texts on Weibo are not as commonly used in the task of author profiling. This is likely due to the centralisation of Weibo in the Chinese population of mainland China, limiting its usage to predominantly China Nationals. Studies done for this platform have used bots, machine learning algorithms to identify authors' age and gender. Data is acquired from Weibo microblog posts of willing participants to be analysed, and used to train algorithms that build concept-based profiles of users to a certain accuracy.
Chat logs