Fenomen Baten'kova i problema verifikatsii avtorstva: mnogomernyy statisticheskiy podkhod k nereshennomu voprosu

0539820 - ÚČL 2021 RIV EE rus J - Journal Article
Šeļa, A. - Plecháč, Petr - Zelenkov, Y.
Fenomen Baten'kova i problema verifikatsii avtorstva: mnogomernyy statisticheskiy podkhod k nereshennomu voprosu.
[The Case of Batenkov and the Problem of Authorship Verification: Multivariate Approach to an Unsolved Question.]
Acta Slavica Estonica. Roč. 12, č. 2020 (2020), s. 131-165. ISSN 2228-2335. E-ISSN 2228-3404
Institutional support: RVO:68378068
Keywords : authorship attribution * stylometry * machine learning * versification * Batenkov * Iliushin * Shapir
OECD category: Specific literatures
Method of publishing: Metadata only

Stat'ya obrashchayetsya k nereshennomu voprosu o poddel'nosti / podlinnosti ryada stikhotvoreniy G. S. Baten'kova v izdanii 1978 g. i predlagayet resheniye, osnovannoye na mnogomernoy stilometrii i mashinnom obuchenii. Avtory demonstriruyut obshchuyu effektivnost' atributsii na osnove «smeshannykh» priznakov teksta (leksika, morfologiya i stikhovaya forma) dlya russkoy poezii pervoy poloviny XIX v. Sluchay Baten'kova formuliruyetsya kak «problema verifikatsii avtorstva» i korpus Psevdo-Baten'kova testiruyetsya na blizost' k originalu s pomoshch'yu metodiki «razoblacheniya». Rezul'taty proverki pokazyvayut, chto somnitel'nyye teksty pri klassifikatsii vedut sebya otnositel'no originala kak teksty «drugogo avtora»: klassifikatoru tak zhe legko otlichit' Baten'kova ot Psevdo-Baten'kova, kak Baten'kova ot drugikh poetov. Etot effekt takzhe nel'zya obyasnit' khronologicheskim nesootvetstviyem dvukh korpusov. Krome togo, teksty Psevdo-Baten'kova v otdel'nom eksperimente demonstriruyut stilisticheskuyu blizost' k poemam, pripisyvayushchimsya A. A. Ilyushinu. Eto pozvolyayet skorrektirovat' vyvody predshestvuyushchego issledovaniya M. I. Shapira i vernut'sya k voprosu o stile i o granitsakh tochnykh metodov v literaturovedenii.

Article follows an unsolved case of possible forgery of G. S. Batenkov’s (1793–1863) poems which were included in the late 1978 collection. A solution is offered that is based on multivariate stylometry and machine learning. Authors demonstrate general effectiveness of authorship attribution in Russian poetry of the first half of 19th century when using mixed features (words, morphology and verse form). The case of Batenkov is formulated as “authorship verification problem” and the corpus of Pseudo-Batenkov is tested against the original using the “unmasking” technique. Results show that dubious texts behave as texts of “other author” in classification task: it is equally easy for a classifier to distinguish between Batenkov and Pseudo-Batenkov as between Batenkov and other poets. Further tests suggest that this effect could not be explained by chronological disrepancy between corpora. In the last attribution experiment Pseudo-Batenkov exhibits a strong stylistic similarity to the poems, presumably written by Iliushin himself. This allows to reconsider some of the conclusions of preceding study by M. I. Shapir and reshape discussion on style and quantitative methods.
Permanent Link: http://hdl.handle.net/11104/0317523