2.1 Author profiling

Computational author profiling consists of inferring author demographics from text. Gender and age recognition are by far the most popular tasks of this kind found in the literature and are often addressed by using supervised machine learning methods. Author profiling has been the centre of a number of shared tasks in the PAN-CLEF series (Rangel and Rosso Reference Rangel and Rosso2019), most notably focused on age and gender prediction in the Twitter domain, although other tasks (e.g., recognising personality traits, language variation, bots, etc.), languages (e.g., Arabic, Dutch etc.) and modalities (e.g., learning from both images and texts) have been addressed as well.

As a brief introduction to recent approaches to author gender and age profiling, Table 1 summarises a number of selected studies in the field, including some of the top-performing systems at PAN-CLEF in 2017 (Basile et al. Reference Basile, Dwyer, Medvedeva, Rawee, Haagsma and Nissim2017), 2018 (Takahashi et al. Reference Takahashi, Tahara, Nagatani, Miura, Taniguchi and Ohkuma2018) and 2019 (Pizarro Reference Pizarro2019).

Table 1. Selected recent approaches to gender and age author profiling according to task (A=age and G=gender), domain, language (En=English, Sp=Spanish, Pt=Portuguese, Ar=Arabic, Fr=French, Ru=Russian, Sw=Swedish), text features (w=word, c=character, or p=part-of-speech n-grams, LIWC (Language Inquiry and Word Count) counts, w2v=Word2vec word embeddings and method (SVM=support vector machines, LR=logistic regression, NB=Naive Bayes, RF=Random Forest, CNN=Convolutional Neural Networks, MLP=multilayer perceptron, LSTM=Long Short-term Memory networks.)

We notice that most approaches are based on Twitter data, make use of word- and character n-gram models, and often based on SVM or logistic regression classifiers. Further details are discussed as follows.

The early work in Nguyen et al. (Reference Nguyen, Trieschnigg, Dogruoz, Gravel, Theune, Meder and de Jong2014) introduces a number of useful insights in gender and age prediction alike. The study compares machine and human performance in gender and age prediction from Twitter texts and discusses a number of the limitations of popular computational approaches to these tasks. The study points out differences between social and biological identities, and shows that, for over 10% of Twitter users, there is a mismatch between their biological sex and the kind of language they use on social media, and that older users tend to be perceived to be younger than what they actually are. The study makes use of Dutch tweets translated to English and compares standard computational models (linear regression for age prediction, and logistic regression for gender prediction) with human evaluation. A majority-vote model obtains an accuracy of 0.84, which is similar to existing author profiling classifiers for English Twitter data.

The work in Basile et al. (Reference Basile, Dwyer, Medvedeva, Rawee, Haagsma and Nissim2017) may be seen as a standard approach to author gender profiling, and it was the overall best-performing participant in the PAN-CLEF-2017 author profiling shared task (Rangel et al. Reference Rangel, Rosso, Potthast and Stein2017). The system obtained 0.83 average accuracy in author gender classification by making use of a linear SVM model with word unigrams and 3.5 character n-gram counts as learning features. Other language- and domain-related features such as part-of-speech (POS) tags and Twitter handles were found to actually harm overall accuracy.

In the work in Reddy et al. (Reference Reddy, Vardhan and Reddy2017), by contrast, the use of POS information plays a more prominent role in a gender classification task. The study introduces a TF-IDF (term frequency–inverse document frequency) weighted POS n-gram model that outperforms a number of standard baseline alternatives (e.g., bag of words, etc.) in the hotel reviews domain.

Unlike most data-driven approaches to author profiling, in Isbister et al. (Reference Isbister, Kaati and Cohen2017), author gender classification is addressed with the aid of psycholinguistic features computed from the Language Inquiry and Word Count (LIWC) dictionary (Pennebaker, Francis, and Booth Reference Pennebaker, Francis and Booth2001). Results from SVM classifiers highlight the role of different LIWC categories in the task and differences across languages.

The work in Kim et al. (Reference Kim, Xu, Qu, Wan and Paris2017) addresses the issues of gender, age and user type Twitter profiling in the English language by classifying graph vertices with the aid of recursive neural networks (RNNs.) To this end, network, text and label information are combined into tree structures and fed into individual RNNs. The approach is found to outperform a number of robust baseline systems (lexica, logistic regression, label propagation, text-associated DeepWalk and Tri-Party Deep Network Representations) in the three tasks under consideration.

The work in Takahashi et al. (Reference Takahashi, Tahara, Nagatani, Miura, Taniguchi and Ohkuma2018) was the overall best-performing system in the PAN-CLEF 2018 author profiling shared task (Rangel et al. Reference Rangel, Rosso, Montes-y-Gómez, Potthast and Stein2018), addressing the issue of gender classification based on multimodal input, that is, conveying both text and image data. To this end, a neural approach called ‘Text Image Fusion Neural Network’ (TIFNN) is introduced in order to leverage both data sources and produce gender predictions accordingly.

Finally, we notice that many of the early approaches to gender and age profiling have been recently outperformed by the work in Rangel et al. (Reference Rangel, Rosso, Zaghouani and Charfi2020), which enhances previous methods with the use of a novel text representation – called LDSE (Low-Dimensionality Statistical Embedding) – that takes into account the word distributions in each author profiling class. In the present work, however, since author profiling is viewed simply as a tool to improve our main task (i.e., authorship attribution), we shall focus on a simple approach to author profiling along the lines in Basile et al. (Reference Basile, Dwyer, Medvedeva, Rawee, Haagsma and Nissim2017), Reddy et al. (Reference Reddy, Vardhan and Reddy2017) and others by making use of word-based models and logistic regression as discussed in Section 4.

2.2 Authorship attribution

Closed-set authorship attribution (hereby called authorship attribution, for short) concerns the computational task of selecting the author of a given document from a well-defined set of candidates (Stamatatos Reference Stamatatos2017). As in the case of author profiling, authorship attribution often resorts to supervised machine learning methods, and it has been the focus of several shared tasks in the PAN-CLEF series (Kestemont et al. Reference Kestemont, Stamatatos, Manjavacas, Daelemans, Potthast and Stein2019), in addition to the related tasks of author clustering (Potthast et al. Reference Potthast, Rangel, Tschuggnall, Stamatatos, Rosso and Stein2017), open-set authorship attribution (Kestemont et al. Reference Kestemont, Stamatatos, Manjavacas, Daelemans, Potthast and Stein2019) and others.

Table 2 summarises a number of recent studies in closed-set authorship attribution. The list, which is by no means complete, is solely intended to illustrate a variety of recent approaches to the task.

Table 2. Selected recent approaches to closed-set authorship attribution according to domain, language (En=English, Sp=Spanish, Fr=French, It=Italian, Pl=Polish), text features (w=word, c=character, or p=part-of-speech n-grams; Glove word embeddings, or PCFG=probabilistic context-free grammars) and methods (SVM=support vector machines, LR=logistic regression, NB=Naive Bayes, RF=Random Forest, Markov Chains, fastText, LDA=Latent Dirichlet allocation, Stacked Ensembles, LSTM=Long short-term memory networks, Convolutional Neural Networks (CNN) and Word Similarity)

Generally speaking, existing approaches to authorship attribution are largely based on word- and character n-gram models, with some methods (Stamatatos Reference Stamatatos2017; Markov et al. Reference Markov, Stamatatos and Sidorov2017) resorting to text distortion (Granados et al. Reference Granados, Cebrián, Camacho and de Borja Rodrguez2011) to omit certain parts of the input text whilst focusing on others. SVM classifiers are among the most popular strategies, and input size – which may be a particular concern in the Twitter domain – has been found to be correlated with overall accuracy (Rocha et al. Reference Rocha, Scheirer, Forstall, Cavalcante, Theophilo, Shen, Carvalho and Stamatatos2017). Most recent studies are devoted to the English language, with the exception of those related to the PAN-CLEF authorship attribution task in Kestemont et al. (Reference Kestemont, Tschugnall, Stamatatos, Daelemans, Specht, Stein and Potthast2018). Individual details are discussed as follows.

The study in Hinh et al. (Reference Hinh, Shin and Taylor2016) makes use of frame semantics from FrameNet (Baker, Fillmore, and Lowe Reference Baker, Fillmore and Lowe1998) to build a bag of frames representation intended to capture an author’s writing style. The model comprises features representing frame semantics statistics such as frame element (FE) counts, average number of FEs per frame and others, and it is compared against a baseline model comprising text-related features such as vocabulary size, word and character counts. Results from a SVM classifier show that the frame semantics authorship attribution model is consistently superior to the baseline in a corpus of adversarial stylometric data.

The work in Stamatatos (Reference Stamatatos2017) makes use of a text distortion method inspired from Granados et al. (Reference Granados, Cebrián, Camacho and de Borja Rodrguez2011), in which rare words are replaced by sequences of a special symbol ‘*’, and more frequent words are kept unchanged. In doing so, authorship attribution SVM classifiers are able to focus on the text fragments that are deemed more relevant to the task. Results suggest, among other findings, that the method does improve overall accuracy and that function words are less suitable to text distortion.

Text distortion is also performed at the pre-processing stage of input texts for authorship attribution in Markov et al. (Reference Markov, Stamatatos and Sidorov2017). In this case, numbers, named entities and highly frequent words are replaced by special symbols. Results from SVM and multinomial Naive Bayes classification suggest that the method compares favourably to a standard bag-of-words approach without pre-processing.

The work in Shrestha et al. (Reference Shrestha, Sierra, Gonzalez, Rosso, Montes-Y-Gomez and Solorio2017) investigates a number of CNN architectures for authorship attribution in social media texts, taking as an input character unigram and bigram embeddings, and skip-gram word embedding representations. Results are compared against those obtained by a range of baseline systems, including the use of logistic regression with variable length character n-grams, and Long Short-Term Memory networks (LSTMs) with character bigrams.

The work in Rocha et al. (Reference Rocha, Scheirer, Forstall, Cavalcante, Theophilo, Shen, Carvalho and Stamatatos2017) focuses on the issues of closed- and open-set authorship attribution (of which only closed-set scenarios are presently dealt with) with limited input data using short texts from a corpus of 10 million tweets posted by 10,000 users (authors). The work makes use of SVM, Random Forest, distance-based and text compression methods built from word, character and POS n-grams. A number of experiments were carried out by varying both the number of candidate authors under consideration and the number of input texts per author. Among other findings, the study suggests that the text compression method outperforms the alternatives for small input sizes and that overall accuracy decreases linearly as the number of candidate authors increases.

The study at Sundararajan and Woodard (Reference Sundararajan and Woodard2018) investigates the role of syntax and word choice in authorship attribution, which may be particularly relevant to cross-genre scenarios in which content-based information does not play a significant role. Syntax is investigated with the aid of context-free probabilistic grammars (PCFG) and Markov chain models, and the issue of word choice is addressed by masking out certain words or topics (which may be seen as an instance of text distortion) corresponding to different POS categories. Results suggest that cross-genre scenarios may benefit from syntactic knowledge, whereas both single- and cross-domain scenarios may benefit from lexical knowledge. Moreover, purely syntactic models were found to be insufficient by themselves, and may require combination with more content-oriented (e.g., character-based) models. In particular, common nouns, verbs, adjectives and adverbs were found to help author identification, whereas proper nouns do not.

The study in Stevenson et al. (Reference Stevenson, Vlachos and Sari2017) addresses the use of continuous word and character n-gram representations for authorship attribution in four domains using fastText (Joulin et al. Reference Joulin, Grave, Bojanowski and Mikolov2017). In doing so, the model focuses on short word and character sequences, but it does not keep track of longer dependencies. Feature representations and classifiers are built jointly by adapting the fastText shallow architecture, and results suggest that the use of continuous character n-gram representations outperform a number of baseline systems (and the use of continuous word n-grams) in two domains (news and reviews.) On the other hand, the use of topic modelling was still superior in the case of legal texts authorship attribution.

The work in Patchala and Bhatnagar (Reference Patchala and Bhatnagar2018) introduces an authorship attribution model based on topic-independent syntactic templates built from each candidate author of interest, and which are intended to represent an individual’s writing style. Results obtained from a number of standard classifiers (e.g., SVM, Naive Bayes and others) suggest that the combination of parsed tree structures and additional syntactic features outperforms the use of individual features alone.

The study in Reddy et al. (Reference Reddy, Reddy, Chand and Venkannababu2018) introduces an instance-based authorship attribution method that relies on author-specific document weights to represent input texts, rather than document features or terms. Document weights are obtained by first computing terms weights, which are subsequently normalised by author. A number of experiments – with and without document weighting – were carried out using standard classifiers (logistic regression, Naive Bayes and Random Forest) based on a bag of words model. Results from a small (10-authors) reviews corpus suggest that document weighting generally increases task accuracy.

The work in Jafariakinabad and Hua (Reference Jafariakinabad and Hua2019) presents a neural model that encodes document information from lexical, syntactic and structural levels for authorship attribution. In this approach, syntactic and lexical sentence representations are jointly encoded, and subsequently an attention-based hierarchical network encodes the syntactic and semantic structures of input texts themselves while rewarding those that help capturing the writing style of their authors. The model is evaluated against a number of SVM and CNN baseline systems, including the approaches in Shrestha et al. (Reference Shrestha, Sierra, Gonzalez, Rosso, Montes-Y-Gomez and Solorio2017) and Stevenson et al. (Reference Stevenson, Vlachos and Sari2017). Results show the strength of each individual level of document information and suggest that the proposed model outperforms the baseline alternatives and its individual components alike.

Unlike existing machine learning approaches to authorship attribution, the work in Sharon Belvisi et al. (Reference Sharon Belvisi, Muhammad and Alonso-Fernandez2020) takes a forensic approach to the task by comparing the use of standard n-gram and stylometric features (e.g., character, word and punctuation counts etc.) through text similarity. More specifically, the evaluation of different features is carried out by measuring the similarity between representations of different authors using Cosine, Euclidean and Manhattan distances. Results based on a small (40-users) Twitter corpus suggest that the use of idiosyncratic features (e.g., misspellings, abbreviations, emoji counts, etc.) outperforms the use of n-gram counts by a small margin.

Finally, the stack ensemble approach in Custódio and Paraboni (Reference Custódio and Paraboni2019) will be taken as the starting point to the present work, and for that reason is discussed in more detail in the next section.

2.3 EACH-USP ensemble approach to authorship attribution

The EACH-USP approach to closed-set authorship attribution described in Custódio and Paraboni (Reference Custódio and Paraboni2019) is based on the assumption that identifying the author of a given document may require relying on multiple knowledge sources. To this end, the approach makes use of standard word- and character-based n-gram models, and an additional character-based model subject to text distortion (Granados et al. Reference Granados, Cebrián, Camacho and de Borja Rodrguez2011). The output probabilities of the three models – hereby called word, char and distorted char – are combined in a stack architecture (Wolpert Reference Wolpert1992) and subject to a second-level logistic regression classifier to determine the author of an input document. This architecture is illustrated in Figure 1.

Figure 1. EACH-USP ensemble authorship attribution, adapted from Custódio and Paraboni (Reference Custódio and Paraboni2019). Input documents (left) are split into train and test portions, and pre-trained word, char and distorted char models (centre) are built from training data. Test documents are submitted to the individual classifiers independently, and their predictions are taken as the input to the final, second-level classifier (right).

Text distortion has been introduced in Granados et al. (Reference Granados, Cebrián, Camacho and de Borja Rodrguez2011) and has been previously considered in authorship attribution (Stamatatos Reference Stamatatos2017), deception detection (Sánchez-Junquera et al. Reference Sánchez-Junquera, nor Pineda, y Gómez, Rosso and Stamatatos2020) and other tasks, and it is largely intended to mask out words that are not relevant to the task. In Custódio and Paraboni (Reference Custódio and Paraboni2019), by contrast, text distortion is performed at the character level. More specifically, the model replaces every character in the input text – except punctuation and diacritics – for a ‘*’ symbol so that the model is able to focus on these particular patterns. An example of text distortion of this kind – rendered in Portuguese to show diacritics usage – is illustrated in Table 3.

Table 3. An example of text distortion in the EACH-USP approach (in Portuguese)

The attention to punctuation and diacritics patterns has been found to be particularly useful for more general, cross-domain authorship attribution tasks in multiple languages, as in Kestemont et al. (Reference Kestemont, Tschugnall, Stamatatos, Daelemans, Specht, Stein and Potthast2018). As for the combination of the three individual classifier components, this works as follows. First, the set of d input documents is vectorised by making use of a word-, char- or distortion-based feature extraction function V(d) as required by each model (or channel), and the resulting feature set X is normalised by a function N(X). Next, X is subject to PCA dimensionality reduction, and a multinomial classifier generates the probability $P(Y=k)$ for each class k.

First-level classifiers are optimised by making use of a second-level model $\sigma (\sum_c \sum_i(w_{ci} * c_i) + k) $ , where $c_i$ is the probability of a candidate author i being the actual author of the given document according to the c classifier, $w_{ci}$ is the weight of $c_i$ , k is a constant and $\sigma$ is the sigmoid function. This produces a new s vector of i probabilities of each candidate author (or class) being the author of the document.

This stack ensemble approach was evaluated at the PAN-CLEF authorship attribution shared task in Kestemont et al. (Reference Kestemont, Tschugnall, Stamatatos, Daelemans, Specht, Stein and Potthast2018), which considered cross-domain authorship attribution scenarios based on fan fiction texts written in five languages, and required identifying the author of a text written in a particular genre (e.g., Harry Potter) based on texts written in a different genre (e.g., Star Wars.) Given the top-performing results reported in Kestemont et al. (Reference Kestemont, Tschugnall, Stamatatos, Daelemans, Specht, Stein and Potthast2018), and the observation that a stack architecture of this kind may be easily extended with any number of additional (e.g., author profiling) classifiers, this approach will be taken as the basis to the present work as well.

3.1 Overview

The present analysis makes use of the b5-corpus of Facebook texts described in Ramos et al. (Reference Ramos, Neto, Silva, Monteiro, Paraboni and Dias2018), which will be further discussed in Section 4.1 as part of our main author profiling experiments. In this experiment, binary gender labels (male/female) available from the corpus are added as a fourth information source to the authorship attribution ensemble in Custódio and Paraboni (Reference Custódio and Paraboni2019), that is, in addition to the word, character and text distortion channels described in the previous Section 2.3. The resulting ensemble is illustrated in Figure 1 and essentially differs from the original architecture only by presenting a fourth (blue) channel at the bottom, which is intended to represent the gender label information taken from the input texts.

The use of binary gender information in this way is similar to the ap.label approach to be discussed in Section 4.2. In its present form, 0/1 gender labels are combined with the probabilities obtained by the three ensemble components and taken as the input to the second-level authorship attribution stack classifier.

Using gender labels available from the input text will provide us with an upper limit for the accuracy that the ensemble authorship attribution model may be able to achieve when using an optimal author profiling classifier. In practice, however, author profiling classifiers will most likely obtain much lower results. To shed light on this issue, we ran a number of simulations in which different levels of noise were added to the model, so that the actual gender information was corrupted by a certain margin. By comparing multiple authorship attribution scenarios based on gender estimates of varying degrees of robustness, we would like to establish the lower limit of accuracy that a gender classifier would be expected to achieve in this particular setting.

3.2 Procedure

The 50 authors with the largest amount of text available from the corpus were selected for this analysis, and their texts were split into document units (or posts) at line breaks. The study consisted of comparing authorship attribution results obtained by the standard EACH-USP approach in Custódio and Paraboni (Reference Custódio and Paraboni2019), which is presently taken as a baseline system, and its extended version that includes gender information with a certain level of added noise. Assuming that gender labels available from the corpus are 100% correct, we tested a number of scenarios in which gender information was corrupted so as to obtain 90%, 80%, 70%, 60% and 50% accuracy, hence simulating author profiling classifiers of different levels of robustness.

Testing was carried out as follows. First, two authors are randomly selected and taken as the input to both models (with and without gender information.) Next, additional authors are randomly selected one at a time, and the procedure is repeated until reaching 50 authors. For the largest (i.e., 20-author) setting, this corresponds to 5600 train and 2400 test documents. At each turn, we compute accuracy, precision, recall and F1 measures. In order to minimise possible effects of random selection, the experiment is repeated 20 times, and we report its overall mean results.

4.1 Data

Using author profiling classifiers to aid authorship attribution requires text documents labelled with both author demographics information (in order to train the author profiling classifiers) and unique author identifiers (to train the authorship attribution model proper.) This unfortunately rules out many of the existing corpora available for the purpose of authorship attribution, including those made available by the PAN-CLEF shared tasks (Kestemont et al. Reference Kestemont, Stamatatos, Manjavacas, Daelemans, Potthast and Stein2019) since those corpora are generally labelled only with author identifiers, but not with author demographics. Corpora developed for author profiling tasks, on the other hand, will obviously provide demographics information, but author identifiers are often unavailable.

Based on these observations, we selected a number of publicly available corpora in different domains and languages, and whose text documents are suitably labelled for both author profiling and authorship attribution tasks as required by our combined approach. More specifically, our models will be built from text in four domains: blog texts from the Blog Authorship corpus (Schler et al. Reference Schler, Koppel, Argamon and Pennebaker2006), Facebook posts from the b5-post corpus (Ramos et al. Reference Ramos, Neto, Silva, Monteiro, Paraboni and Dias2018), short essay texts about topics of a moral nature (e.g., abortion legalisation, death penalty, etc.) from the BRmoral corpus (dos Santos and Paraboni Reference dos Santos and Paraboni2019; Pavan et al. Reference Pavan, dos Santos, Lan, ao Trevisan Martins, dos Santos, Deutsch, da Costa, Hsieh and Paraboni2020) and Twitter data from the TwiSty corpus (Verhoeven, Daelemans, and Plank Reference Verhoeven, Daelemans and Plank2016). We notice that, in addition to providing a certain level of variety to our experiments (and hence reducing possible effects of topical bias and others, as discussed in Sari, Stevenson, and Vlachos Reference Sari, Stevenson and Vlachos2018), some of these choices were motivated by our particular interest in Portuguese NLP, or were aimed at investigating the use author profiling tasks beyond standard gender and age classification.

All data sets are labelled with unique author identifiers. Blogs, Facebook and essay texts are labelled so as to support multiple author profiling tasks in one single language each, whereas Twitter texts are labelled only with author gender information available in six languages (and which may therefore be regarded as six independent corpora.) Descriptive statistics for each corpus are summarised in Table 5 and further discussed below.

Table 5. Corpus descriptive statistics representing the number of authors, the text unity taken as an input document, the number of text unities in each corpus, their average number of words and author profiling tasks supported by the existing labels (G=gender, A=age bracket, I=IT background, E=level of education, P=political orientation, R=degree of religiosity.)

Possible author profiling tasks are determined by the labels available from each corpus. All corpora are labelled with author binary gender (G) information, and therefore support binary (male/female) gender classification. With the exception of the Twitter domain, all corpora are also labelled with age (A) information, which has been presently modelled as a 3-class problem based on the distribution of each corpus. IT background is a binary label available from the b5-post (Facebook) and BRmoral (Essay) corpora only, both of which containing a significant proportion of text produced by students in Computer Science and related fields, and which indicates whether each author in the corpus has this kind of background or not. Level of Education (E), political orientation (P) and degree of religiosity (R) are crowd-sourced, self-reported labels available in the essay domain represented by the BRmoral corpus. Each of these labels supports a ternary classification problems (from basic to superior education, from left to right political orientation and from no religious at all to highly religious.) For details regarding the BRmoral corpus and its annotation scheme, we refer to Pavan et al. (Reference Pavan, dos Santos, Lan, ao Trevisan Martins, dos Santos, Deutsch, da Costa, Hsieh and Paraboni2020).

4.2 Author profiling and authorship attribution models

As in the present work, author profiling classifiers have been developed only as a support to the main task of authorship attribution, in what follows we take a standard approach to the task by making use of TF-IDF unigram counts and multinomial logistic regression with univariate feature selection along the lines of Hsieh, Dias, and Paraboni (Reference Hsieh, Dias and Paraboni2018) and others. In all models, logistic regression uses L2 regularisation and newton-cg solver with a 0.0001 tolerance as a stopping criteria. For reasons discussed below, depending on the authorship attribution strategy under consideration, author profiling predictions obtained by performing logistic regression may be taken either as class probabilities or as actual class labels.

Regarding the authorship attribution task proper, the present work essentially extends the EACH-USP stack authorship attribution approach in Custódio and Paraboni (Reference Custódio and Paraboni2019) by adding author profiling classifiers to the existing ensemble of word, character and distorted character models as discussed in Section 2. In other words, the actual architecture is similar to the previous Figure 2, except that author demographics information will be presently inferred from text automatically with the aid of author profiling classifiers, rather than taken from ground truth corpus labels. This is illustrated in Figure 4 using two author profiling modules as an example (in light colour, at the bottom) and further discussed below.

Figure 4. Ensemble authorship attribution with added author profiling classifiers. For the main authorship attribution task, 20 input documents labelled with author ids only (left) are split into train and test data, and word, char and distorted char models (top middle) are built from this training data. From the reminder of the data and accompanying demographics labels (disjoint data, in the bottom left), the auxiliary author profiling classifiers are built. Test documents are submitted to each individual classifier independently, and their predictions are taken as the input to the final, second-level classifier (right).

Two strategies for adding author profiling predictions to the ensemble, hereby called ap.prob and ap.label, are presently considered. These strategies differ from each other only in the way their output predictions are represented. In ap.prob, we use author profiling predictions represented as probabilities not unlike the output of any of the existing components of the stack ensemble model. In ap.label, by contrast, we use class labels predictions (e.g., for gender, age etc.) In doing so, we would like to investigate the extent to which the present authorship attribution tasks may benefit from having access to more fine-grained probabilities or more coarse-grained class label predictions.

The use of author profiling probabilities from a given input document in ap.prob is illustrated as follows. Let us consider, for instance, a gender classifier that predicts that the author of the document has a 0.43 probability of being male, and hence a 0.57 probability of being female. In a binary classification task of this kind, both probabilities are taken as an input to the second-level classifier (that is, in addition to the existing probabilities predicted by the original word, char and distorted char classifiers.) Similarly, for ternary author profiling classes (e.g., education level etc.), the three probabilities are considered.

An example of how the input to the second-level classifier is represented in ap.prob is illustrated in Table 6, in which probabilities provided by the Word, Char (character) and text distortion (Dist) modules of the EACH-USP ensemble for an individual candidate author are appended to his/her gender probabilities. This creates a set of five probabilities associated with each author, which are to be submitted to the second-level authorship attribution classifier.

Table 6. Second-level classifier input represented as probabilities in ap.prob

Regarding the use of class label predictions in ap.label, author profiling probabilities are replaced by class labels directly or, more specifically, by assigning the value 1 to the class of highest probability and 0 to all the others. Thus, for instance, the 0.57 probability of being female in the previous example would be replaced by a 1 value, and the 0.43 probability of being male would be replaced by 0. An example of this representation using a ternary class (Education e1.e3) for an individual author is illustrated in Table 7, in which the class of highest probability is assumed to be e2.

Table 7. Second-level classifier input represented as class labels in ap.label

5.1 Train and test sets

Central to the current approach is the separation between data for our main task – authorship attribution – and for the auxiliary author profiling classifiers. Authorship attribution data consist of a set of train and test documents produced by 20 selected authors as discussed below, and it is labelled only with author identifiers. Author profiling data, by contrast, comprise all documents produced by other authors (i.e., those outside the 20-author group), and it is labelled with author demographics only (e.g., gender, age, etc.)

The organisation of each corpus into training and test sets takes into account the differences in granularity of the author profiling and authorship attribution tasks. The present author profiling models take as an input the set of all texts produced by an individual, which are concatenated as a single document labelled with their corresponding demographics. The authorship attribution models, by contrast, require multiple text samples from each candidate author (or else author identification would become trivial) and, as a result, take as an input individual text unities (i.e., Facebook and blog posts, sentences or tweets) as described in Section 4.1.

As the main focus of the present work is the authorship attribution task, we selected from each corpus the 20 authors with the largest volume of text available. These sets of 20 authors are taken to be the test data of our main authorship attribution approach in each domain and will be considered in a number of experiments in multiple test scenarios conveying from 2 to 20 candidates each. The choice for the authors with the largest possible amount of text data is intended to minimise situations in which the baseline approach in Custódio and Paraboni (Reference Custódio and Paraboni2019) may fail due to lack of data, which would have obscured the role of the author profiling classifiers as an aid to the authorship attribution task. The issue of input size in authorship attribution is addressed in detail in Rocha et al. (Reference Rocha, Scheirer, Forstall, Cavalcante, Theophilo, Shen, Carvalho and Stamatatos2017). Test set author profiling class distributions are illustrated in Table 8. We notice, however, that these class labels are not taken into account by the present approach and are presented only as a means to illustrate how author profiling estimates may help authorship attribution.

Table 8. Demographics distribution in the 20-author test set for Gender (female/male), IT background (no/yes), Age, Education, Religiosity and Politics levels (low/medium/high)

Test sets selected from each domain are naturally more balanced towards some classes than others and, as the comparison among author profiling classifiers (e.g., gender, age, etc.) will require a fixed test set for each domain, class imbalance may impact the results of the present authorship attribution approach. For instance, we notice using a gender classifier is arguably less helpful if, for example, most test authors turn out to be of the same gender. Keeping balanced test sets for all author profiling classes is, however, impractical for a number of reasons. First, we notice that this would require a large number of distinct author profiles to cover all possible class values (e.g., the Essay domain would require a test set consisting of at least 324 distinct authors selected out of a corpus containing only 510 individuals). Moreover, many profiles are considerably rare, or simply do not occur at all in the data (e.g., there are relatively few individuals who belong to the more extreme classes of education, politics and religiosity; most IT people tend to be male, etc.)

In order to minimise these difficulties, in the evaluation of our authorship attribution approach we will keep the natural author profiling class imbalance as is, and we will resort instead to multiple random tests as discussed in Section 5.3. The issue of class imbalance will also be revisited in the light of our results as discussed in Section 6.3.

Finally, leaving the 20 test authors aside, the remaining portion of each corpus is concatenated (i.e., disregarding author identifiers) and taken as training data for the auxiliary author profiling classifiers in each domain. Thus, the training data for author profiling does not include any text produced by the members of the 20-author group and, conversely, authorship attribution does not rely on the actual demographics information associated with the authors under identification, making instead its own predictions based on a disjoint data set.

The number of train and test documents (i.e., text units, cf. previous Table 5) for each task are summarised in Table 9, based on the largest possible (i.e., 20-author) evaluation setting.

Table 9. Train and test instances for the author profiling and authorship attribution tasks in the 20-author evaluation setting

5.2 Author profiling evaluation

Prior to the evaluation of the present authorship attribution approach, we built and evaluated its individual components, that is, the author profiling classifiers that could be built from the existing labels in each corpus. To this end, we performed univariate feature selection over development data using the ANOVA function and the F1 metrics to obtain the k-best text features (i.e., words) in each domain and language. Optimal values were searched within the 3000–20,000 features range at 1000 intervals and are summarised in Table 10.

Table 10. Author profiling univariate feature selection k-best values

Evaluation of the author profiling models was carried out by performing 10-fold cross validation over the training data and by considering a simple majority class baseline for illustration purposes. For all author profiling classifiers, we measure mean precision, recall, F1 and accuracy scores.

5.3 Authorship attribution evaluation

Multiple authorship attribution evaluation experiments were carried out by considering random sets of candidate authors drawn from the 20-author test set. With the exception of the Blog domain, tests were carried out by varying the number of candidate authors from 2 to 20. In the case of blogs, only tests involving 5, 10, 15 and 20 candidates were considered due to computational costs. As a means to obtain a balanced (authorship attribution) classification setting, the number of input texts taken from each candidate author is kept constant within each task by considering the smallest set size of the group.

In order to reduce the possible effects of random selection (e.g., in case of author profiling class imbalance when most authors turn out to belong to the same gender, etc.), evaluation was repeated 20 times by varying the candidates randomly, and also by randomly selecting different train and test documents. More specifically, we performed 20 runs * (3 corpora * 19 non-blog candidate set sizes) + 20 runs * (1 corpus * 4 blog candidate set sizes) experiments, making 1220 randomised authorship attribution evaluation tasks in total. Given the large number of evaluation scenarios, in what follows we will report overall mean results only.

The two versions of our current approach – ap.prob and ap.label – are to be compared against the standard EACH-USP ensemble baseline system in Custódio and Paraboni (Reference Custódio and Paraboni2019) whilst measuring mean accuracy scores for all models. In doing so, our goal is to verify whether using author profiling classifiers improves results over the original approach that does not have access to author demographics predictions.