Unsupervised Aspect-Level Sentiment Controllable Style Transfer

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 303–312 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 303 Unsupervised Aspect-Level Sentiment Controllable Style Transfer Mukuntha N S§, Zishan Ahmad§, Asif Ekbal, Pushpak Bhattacharyya Department of Computer Science and Engineering Indian Institute of Technology Patna Bihar, India {mukuntha.cs16,1821cs18,asif,pb}@iitp.ac.in Abstract Unsupervised style transfer in text has pre- viously been explored through the sentiment transfer task. The task entails inverting the overall sentiment polarity in a given input sen- tence, while preserving its content. From the Aspect-Based Sentiment Analysis (ABSA) task, we know that multiple sentiment polar- ities can often be present together in a sen- tence with multiple aspects. In this paper, the task of aspect-level sentiment controllable style transfer is introduced, where each of the aspect-level sentiments can individually be controlled at the output. To achieve this goal, a BERT-based encoder-decoder archi- tecture with saliency weighted polarity injec- tion is proposed, with unsupervised training strategies, such as ABSA masked-language- modelling. Through both automatic and man- ual evaluation, we show that the system is suc- cessful in controlling aspect-level sentiments. 1 Introduction With a rapid increase in the quality of generated text, due to the rise of neural text generation mod- els (Kalchbrenner and Blunsom, 2013; Cho et al., 2014; Sutskever et al., 2014; Vaswani et al., 2017), controllable text generation is quickly becoming the next frontier in the field of text generation. Con- trollable text generation is the task of generating re- alistic sentences whose attributes can be controlled. The attributes to control can be: (i). Stylistic: Like politeness, sentiment, formality etc, (ii). Content: Like information, entities, keywords etc. or (iii). Ordering: Like ordering of information, events, plots etc. Controlling sentence level polarity has been well explored as a style transfer task. Zhang et al. (2018) used unsupervised machine translation techniques for polarity transfer in sentences. Yang et al. (2018) §equal contribution The service was speedy and the salads were great, but the chicken was bland and stale. The service was slow, but the salads were great and the chicken was tasty and fresh. Service - Positive Salads - Positive Chicken - Negative Service - Negative Salads - Positive Chicken - Positive Query: Service - Negative Salads - Positive Chicken - Positive Figure 1: An example of the proposed aspect-level sen- timent style transfer task used language models as discriminators to achieve style (polarity) transfer in sentences. Li et al. (2018a) proposed a simpler method where they deleted the attribute markers and devise a method to replace or generate the target attribute-key phrases in the sentence. In this paper we explore a more fine-grained style transfer task, where each aspect’s polarities can be changed individually. Recent interest in Aspect-Based Sentiment Analysis (ABSA) (Pontiki et al., 2014) has shown that sentiment information can vary within a sentence, with differing senti- ments expressed towards different aspect terms of target entities (e.g. ‘food’, ‘service’ …

Content truncated to 500 words