Learning Stylometric Representations for Authorship Analysis

Abstract

Authorship analysis (AA) is the study of unveiling the hidden properties of authors from a body of exponentially exploding textual data. It extracts an author's identity and sociolinguistic characteristics based on the reflected writing styles in the text. It is an essential process for various areas, such as cybercrime investigation, psycholinguistics, political socialization, etc. However, most of the previous techniques critically depend on the manual feature engineering process. Consequently, the choice of feature set has been shown to be scenario-or dataset-dependent. In this paper, to mimic the human sentence composition process using a neural network approach, we propose to incorporate different categories of linguistic features into distributed representation of words in order to learn simultaneously the writing style representations based on unlabeled texts for authorship analysis. In particular, the proposed models allow topical, lexical, syntactical, and character-level feature vectors of each document to be extracted as stylometrics. We evaluate the performance of our approach on the problems of authorship characterization and authorship verification with the Twitter, novel, and essay datasets. The experiments suggest that our proposed text representation outperforms the bag-of-lexical-n-grams, Latent Dirichlet Allocation, Latent Semantic Analysis, PVDM, PVDBOW, and word2vec representations.

Improving author verification based on topic modeling

“…We focus on the two most well‐known topic modeling approaches: Latent Semantic Indexing (LSI) and Latent Dirichlet Allocation (LDA) and examine their suitability for each author verification paradigm. It is demonstrated that topic modeling can considerably increase the effectiveness of author verification methods when an appropriate topic modeling technique is selected and is adequately fine‐tuned. We examine the effect of using either a limited set of documents or an enriched document collection to extract latent topics and it is demonstrated that the latter assists author verification methods to further increase their effectiveness. We report experimental results on benchmark data sets developed during the relevant PAN‐2014 and PAN‐2015 shared tasks in author verification that are directly compared with state‐of‐the‐art methods under the same settings. The performance of the methods presented in this study is quite competitive to the best results reported so far for these data sets, demonstrating that topic modeling can be an efficient and effective alternative to more sophisticated methods (for example, based on representation learning, distributed document representation, or neural network language models (Bagnall, ; Ding et al, )) for the author verification task. We examine the effect of genre of external documents when extrinsic author verification methods are combined with topic modeling techniques. It is demonstrated that verification models based on genre‐agnostic external documents are very competitive, but they are outperformed by models using external documents of the same genre with that of the questioned documents.…”

“…This approach has a relatively large number of hyperparameters to be set. The experimental results on PAN‐2014 benchmark data sets demonstrated that this sophisticated method can outperform existing methods as well as baselines based on topic modeling techniques (following an intrinsic and profile‐based paradigm with a prefixed number of latent topics) as well as other distributed word and document representations (word2vec, doc2vec) (Ding et al, ). In this article, we show that similar or even higher performance can be achieved by appropriately combining fine‐tuned topic modeling techniques and author verification paradigms.…”

“…Profile‐based methods are better able to handle cases with limited text length since they concatenate all available documents. This paradigm is more popular in recent studies (Ding et al, ; Halvani et al, ; Potha & Stamatatos, ). There are also some attempts to combine instance‐based and profile‐based paradigms (Bagnall, ; Sari & Stevenson, ).…”

“…Although the main idea is similar to META‐PAN14, this method was not equally effective, since several PAN‐2015 submissions produced very low results affecting the quality of the ensemble. Hernández and Calvo (): This is an intrinsic and instance‐based method that utilizes topic modeling. LDA features are used to represent known and unknown documents and then the difference between them reveals a tendency to use roughly similar or different latent topics. Ding et al (): This is an intrinsic and profile‐based method that uses a joint neural network to learn distributed word representations as well as topical and lexical biases related to the global topic of the document and the personal bias of the author to use specific words given that topic. The reported results of this method are the best ones so far for the PAN‐2014 data sets.…”

“…2. Profile-based paradigm: All available samples in D known are first concatenated in one big document and a single representation vector is extracted (author's profile) (Ding et al, 2019;Halvani et al, 2016;Kocher & Savoy, 2017;Potha & Stamatatos, 2014). The methods in this paradigm follow an author-centric approach where the differences between documents by the same author are disregarded.…”