Inducing Positive Perspectives with Text Reframing

Sentiment transfer is one popular example of a text style transfer task, where the goal is to reverse the sentiment polarity of a text. With a sentiment reversal comes also a reversal in meaning. We introduce a different but related task called positive reframing in which we neutralize a negative point of view and generate a more positive perspective for the author without contradicting the original meaning. Our insistence on meaning preservation makes positive reframing a challenging and semantically rich task. To facilitate rapid progress, we introduce a large-scale benchmark, POSITIVE PSY- CHOLOGY FRAMES, with 8,349 sentence pairs and 12,755 structured annotations to explain positive reframing in terms of six theoreticallymotivated reframing strategies. Then we evaluate a set of state-of-the-art text style transfer models, and conclude by discussing key challenges and directions for future work. To download the data, see https://github. com/GT-SALT/positive-frames

Introduction. Text style transfer (TST) has received much attention from the language technologies community (Hovy, 1987; Jin et al., 2020), where the goal is to change some attribute, like the sentiment of the text, without changing any attribute-independent content (Mir et al., 2019; Fu et al., 2018; Logeswaran et al., 2018). Some TST applications such as de-biasing (Pryzant et al., 2020; Ma et al., 2020) and paraphrasing (den Bercken et al., 2019; Xu et al., 2012) require meaning-preserving transformations, while political leaning (Prabhumoye et al., 2018), sentiment (Shen et al., 2017; Hu et al., 2017), and topical transfer (Huang et al., 2020) allow for a change In this work, we introduce a closely related task— positive reframing—that differs from sentiment TST in important ways. We effectively reframe negative text by inducing a complementary positive viewpoint (e.g. glass-half-full), which nevertheless supports the underlying content of the original sentence.

Discussion / Conclusion. This work introduces a new and challenging NLG task called positive reframing. The objective is to construct a more positive outlook as a way of rephrasing a negative source text such that the meaning of that source is preserved. Our parallel dataset, POSITIVE PSYCHOLOGY FRAMES, will serve as a benchmark that will enable sustained work on this task. We experiment with many of the leading style-transfer models and show that these models can learn to shift from a negative to a more positive perspective using a combination of strategies from positive psychology. Importantly, the best models are fluent and effective reframing systems that can learn to largely preserve the meaning of the original text, even under a perspective shift. However, these models still struggle to generate reasonable positive perspectives, and even the best models are still prone to errors. We discuss four key error classes: insubstantial changes, contradictions to the premise, self-contradictions, and hallucinations, as shown in Error Analyses in Section 5.5. Overall, this suggests that our dataset can serve as a useful benchmark for understanding well-motivated positive reframing strategies and equipping natural language generation systems with positive perspectives.