Delete, Retrieve, Generate: a Simple Approach to Sentiment and Style Transfer

By Juncen Li, Robin Jia, He He, Percy Liang. 2018

586 citations recorded by the source corpus.

We consider the task of text attribute transfer: transforming a sentence to alter a specific attribute (e.g., sentiment) while preserving its attribute-independent content (e.g., “screen is just the right size” to “screen is too small”). Our training data includes only sentences labeled with their attribute (e.g., positive and negative), but not pairs of sentences that only differ in the attributes, so we must learn to disentangle attributes from attribute-independent content in an unsupervised way. Previous work using adversarial methods has struggled to produce high-quality outputs. In this paper, we propose simpler methods motivated by the observation that text attributes are often marked by distinctive phrases (e.g., “too small”). Our strongest method extracts content words by deleting phrases associated with the sentence’s original attribute value, retrieves new phrases associated with the target attribute, and uses a neural model to fluently combine these into a final output. Based on human evaluation, our best method generates grammatical and appropriate responses on 22% more inputs than the best previous system, averaged over three attribute transfer datasets: altering sentiment of reviews on Yelp, altering sentiment of reviews on Amazon, and altering image captions to be more romantic or humorous.

Subfields: Artificial Intelligence, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

Citations in the corpus

Self-Refine: Iterative Refinement with Self-Feedback (2023, background): “While a large body of work on style transfer is directed at sentence-level sentiment transfer (Li et al., 2018; Prabhumoye et al., 2018), we focus on transferring the sentiment of entire reviews, making the task challenging and providing opportunities for iterative improvements.”
A Recipe for Arbitrary Text Style Transfer with Large Language Models (2021, methodology): “Furthermore, following Li et al. (2018) and Sud-hakar et al. (2019), we perform automatic evaluation for sentiment style transfer since there are classiﬁers available for these styles.”
What Makes Good In-Context Examples for GPT-3? (2021, background): “Some representative applications in the ﬁeld of deep learning include machine translation (Gu et al., 2018), sentiment transfer (Li et al., 2018; Guu et al., 2018), QA (Karpukhin et al., 2020; Mao et al., 2020), dialogue generation (Yan et al., 2016; Cai et al., 2018; Song et al., 2016; Pandey et…”
CommonGen: A Constrained Text Generation Challenge for Generative Commonsense Reasoning (2020, background): “, 2019; J Kurisinkel and Chen, 2019), style (Fu et al., 2018; Luo et al., 2019b; Li et al., 2018), topics (Feng et al.”
Unifying Human and Statistical Evaluation for Natural Language Generation (2019, background): “Such models are known to perform well in tasks with complex outputs such as program generation (Hayati et al., 2018; Hashimoto et al., 2018) and style transfer (Li et al., 2018).”
DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts (2021, methodology): “As a second application we consider the well-studied task of controlling the polarity of text’s sentiment (e.g., Li et al., 2018; Sudhakar et al., 2019), steering towards either positive or negative sentiment.”
A Distributional Approach to Controlled Text Generation (2020, methodology): “As metrics, we use sentiment class expectation Eφ(x), the perplexity according to an external GPT2 small architecture as in (Li et al., 2018), and the diversity metrics introduced in section §3.”
Politeness Transfer: A Tag and Generate Approach (2020, methodology): “Compared to prior work, “Delete, Retrieve and Generate” (Li et al., 2018) (referred to as DRG henceforth) and its extension (Sudhakar et al., 2019) are effective methods to generate outputs in the target style while having a relatively high rate of source content preservation.”
Style Transfer Through Back-Translation (2018, background): “Li et al. (2018) first extract words or phrases associated with the original style of the sentence, delete them from the original sentence and then replace them with new phrases associated with the target style.”
Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects (2019, methodology): “As Table 8 shows, models trained on reviews and tips tend to generate generic phrases (such as ‘i love this place’) which often do not include information that helps users to make de-Dataset Diversity-aware NLG Diversity is an important aspect of NLG systems.”
PowerTransformer: Unsupervised Controllable Revision for Biased Language Correction (2020, methodology): “Note that ParaOnly + noBoost is equivalent to a GPT-based encoder-decoder model, similar to seq2seq frameworks commonly used in paraphrasing tasks (Cao et al., 2017; Li et al., 2018b; Prakash et al., 2016).”
Investigating Societal Biases in a Poetry Composition System (2020, methodology): “For the parameters specific to the “Delete, Retrieve, Generate” architecture (Li et al., 2018), we use n-gram attributes and an attribute salience threshold of 10.”
Blank Language Models (2020, methodology): “Inspired by the way humans perform rewriting, we follow a recent line of work in style transfer (Li et al., 2018; Xu et al., 2018; Wu et al., 2019) that adopts a two-step approach: 1.”
Response Generation by Context-aware Prototype Editing (2018, methodology): “The other approach follows a retrieval-then-edit paradigm, that uses a Seq2Seq model to edit a prototype retrieved from a corpus (Guu et al. 2017; Li et al. 2018; Cao et al. 2018).”
Extract and Edit: An Alternative to Back-Translation for Unsupervised Neural Machine Translation (2019, neutral): “…related to the recent work on applying retrieval mechanisms to augment text generation, such as image captioning (Kuznetsova et al., 2013; Mason and Charniak, 2014), dialogue generation (Song et al., 2016; Yan et al., 2016; Wu et al., 2018) and style transfer (Lin et al., 2017; Li et al., 2018).”
Fair Transfer of Multiple Style Attributes in Text (2019, methodology): “Most of the existing style transfer techniques have focussed on individual and granular styles; for example, altering the sentiment of a review from negative to positive style, transferring a tweet from male to female style, and converting the style of text from one author to other [5], [7], [11], [”
TextSETTR: Label-Free Text Style Extraction and Tunable Targeted Restyling (2021, methodology): “To evaluate our approach and better understand the effects of our various design choices, we test on label-free sentiment transfer, using the Amazon reviews dataset of Li et al. (2018). However, as their training split doesn’t indicate which sentences were adjacent in the original reviews, we make u”
Learning to Control the Fine-grained Sentiment for Story Ending Generation (2019, background): “Previous works mainly study on how to generate a coherent, reasonable and diversiﬁed story ending (Li et al., 2018; Guan et al., 2018; Xu et al., 2018).”