Person identification from text and speech genre samples

Abstract

In this paper, we describe experiments conducted on identifying a person using a novel unique correlated corpus of text and audio samples of the person's communication in six genres. The text samples include essays, emails, blogs, and chat. Audio samples were collected from individual interviews and group discussions and then transcribed to text. For each genre, samples were collected for six topics. We show that we can identify the communicant with an accuracy of 71% for six fold cross validation using an average of 22,000 words per individual across the six genres. For person identification in a particular genre (train on five genres, test on one), an average accuracy of 82% is achieved. For identification from topics (train on five topics, test on one), an average accuracy of 94% is achieved. We also report results on identifying a person's communication in a genre using text genres only as well as audio genres only.

Authorship Attribution Using Text Distortion

“…The unmasking method for author verification of long documents based on very frequent word frequencies was successfully tested in cross-topic conditions (Koppel et al, 2007) but Kestemont, et al (2012) found that its reliability was significantly lower in cross-genre conditions. Function words have been found to be effective when topics of the test corpus are excluded from the training corpus (Baayen et al, 2002;Goldstein-Stewart et al, 2009;Menon and Choi, 2011). However, Mikros and Argiri (2007) demonstrated that function word features actually correlate with topic.…”

“…However, Mikros and Argiri (2007) demonstrated that function word features actually correlate with topic. Other types of features found effective in cross-topic and crossgenre authorship attribution are punctuation mark frequencies (Baayen et al, 2002), LIWC features (Goldstein-Stewart et al, 2009), and character n-grams (Stamatatos, 2013). To enhance the performance of attribution models based on character n-gram features, Sapkota et al (2015) define several n-gram categories and then they combine n-grams that correspond to word affixes and punctuation marks.…”

Masking topic‐related information to enhance authorship attribution

“…In all these cases, a topic is mainly characterized by coarse‐grained thematic areas. Finally, some corpora that control topic and genre have been built (by hiring people for writing under controlled conditions) to explore cross‐domain attribution (Baayen, van Halteren, Neijt, & Goldstein‐Stewart et al, ; Tweedie, ). The latter approach is certainly the most reliable, providing a fine‐grained range of topics and genres.…”

“…Madigan et al () demonstrated that part‐of‐speech features are more effective than word unigrams in cross‐topic conditions. Function words have been found to be effective when topics of the test corpus are excluded from the training corpus (Baayen, et al, ; Goldstein‐Stewart et al, ; Menon & Choi, ). However, Mikros and Argiri () demonstrated that function word features actually correlate with topic.…”

“…Intuitively, function words (i.e., prepositions, articles, etc.) seem robust features in cross‐topic attribution, and Goldstein‐Stewart, Winder, and Sabin () as well as Menon and Choi () reported encouraging results. However, Mikros and Argiri () demonstrated that these features are not immune to topic shifts.…”