Do Topics Make a Metaphor? Topic Modeling for Metaphor Identification and Analysis in Russian

Abstract

The paper examines the efficiency of topic models as features for computational identification and conceptual analysis of linguistic metaphor on Russian data. We train topic models using three algorithms (LDA and ARTM -sparse and dense) and evaluate their quality. We compute topic vectors for sentences of a metaphor-annotated Russian corpus and train several classifiers to identify metaphor with these vectors. We compare the performance of the topic modeling classifiers with other state-of-the-art features (lexical, morphosyntactic, semantic coherence, and concreteness-abstractness) and their different combinations to see how topics contribute to metaphor identification. We show that some of the topics are more frequent in metaphoric contexts while others are more characteristic of non-metaphoric sentences, thus constituting topic predictors of metaphoricity, and discuss whether these predictors align with the conceptual mappings attested in literature. We also compare the topical heterogeneity of metaphoric and non-metaphoric contexts in order to test the hypothesis that metaphoric discourse should display greater topical variability due to the presence of Source and Target domains.

Identifikacija metafore in metonimije v jezikovnih korpusih

“…Pri tem je treba za vsako leksikalno enoto ugotoviti, ali se njen konkretni kontekstualni pomen razlikuje od njenega osnovnega pomena. Postopek je s prilagoditvami značilnostim posameznih jezikov sprožil zanimanje za identifikacijo metaforičnih izrazov in metafor v češčini (Pavlas in drugi, 2018), litovščini (Urbonaitė, 2016), madžarščini (Babarzy in Bencze, 2010), poljščini (Risinski in Mahula, 2015), srbščini (Bogetić, 2019) ter za izdelavo korpusov metafor v ruščini (Badryzlova in Lyashevskaya, 2017), hrvaščini (Despot in drugi, 2019) in kitajščini (Lu in Wang, 2017). Vsem navedenim je skupno, da so v korpus zajeta samo pisna besedila, da so označene samo metaforične besede, metafore ali domene preslikave (ne pa tudi metonimične besede, metonimični prenosi oziroma metonimije) in da je postopek identifikacije in označevanja metaforičnih besed prilagojen posebnostim ciljnega jezika.…”

Automated Metaphor Identification in Russian and Its Implications for Metaphor Studies

“…Metaphor identification systems of the first type are presented, for example, in [23,28]; systems of the second type are described in [5,14,25], among others. Technically, the top-down approach is relatively less challenging, because the metaphor identification system is searching for preconceived patterns.…”

“…We train two types of topic models: LDA and ARTM [30] (the latter comes in two types, sparse and dense). The topic models were incorporated with BERT word embeddings by concatenating topic vectors with averaged BERT vectors (see [5] for further detail).…”