T-STAR: Truthful Style Transfer using AMR Graph as Intermediate Representation

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8805 - 8825 December 7-11, 2022 ©2022 Association for Computational Linguistics T-STAR: Truthful Style Transfer using AMR Graph as Intermediate Representation Anubhav Jangra∗ Preksha Nema∗ Aravindan Raghuveer Google Research, India {anubhavjangra,preksh,araghuveer}@google.com Abstract Unavailability of parallel corpora for training text style transfer (TST) models is a very chal- lenging yet common scenario. Also, TST models implicitly need to preserve the content while transforming a source sentence into the target style. To tackle these problems, an in- termediate representation is often constructed that is devoid of style while still preserving the meaning of the source sentence. In this work, we study the usefulness of Abstract Meaning Representation (AMR) graph as the interme- diate style agnostic representation. We posit that semantic notations like AMR are a natu- ral choice for an intermediate representation. Hence, we propose T-STAR: a model com- prising of two components, text-to-AMR en- coder and a AMR-to-text decoder. We pro- pose several modeling improvements to en- hance the style agnosticity of the generated AMR. To the best of our knowledge, T-STAR is the first work that uses AMR as an intermedi- ate representation for TST. With thorough ex- perimental evaluation we show T-STAR signif- icantly outperforms state of the art techniques by achieving on an average 15.2% higher con- tent preservation with negligible loss (∼3%) in style accuracy. Through detailed human evalu- ation with 90, 000 ratings, we also show that T- STAR has upto 50% lesser hallucinations com- pared to state of the art TST models. 1 Introduction A well accepted definition of style refers to the manner (via linguistic elements like word choices, syntactic structures, metaphors) in which semantics of a sentence are expressed (McDonald and Puste- jovsky, 1985; Jin et al., 2020). Text Style Transfer (TST) is the task of rephrasing an input sentence to contain specific stylistic properties without altering the meaning of the sentence (Prabhumoye et al., 2018). We refer the reader to Jin et al. (2020) for a ∗denotes equal contribution detailed survey of approaches towards TST prob- lem formulation, metrics and models. In the practi- cal scenario, that we consider in this paper, where a large corpus of parallel data is not available (Niu and Bansal, 2018; Ma et al., 2020; Wu et al., 2020), two family of approaches have been proposed in the literature (Jin et al., 2020). 1. Disentanglement: Content and Style are disentangled in a latent space and only the style information is varied to trans- form the sentence. 2. Prototype Editing: Style bearing words in the source sentence are replaced with those corresponding to the target style. The sentence may be further re-arranged for fluency and naturalness. Both the above approaches have draw- backs in the way the style agnostic intermediate representation is constructed, described as follows. First, in the disentangling approaches, it is not easy to verify the efficacy of separation between style and content. Recent approaches (Subramanian et al., 2018; Samanta et al., 2021) …

Content truncated to 500 words