Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
91 tokens/sec
GPT-4o
12 tokens/sec
Gemini 2.5 Pro Pro
o3 Pro
5 tokens/sec
GPT-4.1 Pro
15 tokens/sec
DeepSeek R1 via Azure Pro
33 tokens/sec
Gemini 2.5 Flash Deprecated
12 tokens/sec
2000 character limit reached

Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer (1804.06437v1)

Published 17 Apr 2018 in cs.CL

Abstract: 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., changing "screen is just the right size" to "screen is too small"). Our training data includes only sentences labeled with their attribute (e.g., positive or negative), but not pairs of sentences that differ only in their 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. 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.

Citations (535)

Summary

  • The paper proposes a three-phase process that separately handles deletion, retrieval, and generation to perform effective sentiment and style transfer.
  • Its modular methodology achieves competitive performance by preserving semantic content and yielding measurable improvements in BLEU scores and sentiment accuracy.
  • The framework enhances interpretability and reduces computational overhead, offering a practical alternative to complex end-to-end neural models.

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

The paper "Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer" by Juncen Li, Robin Jia, He He, and Percy Liang introduces a novel methodology for sentiment and style transfer in text. This approach is characterized by a three-step process: deletion of sentiment-laden words, retrieval of potential replacement phrases, and generation of new stylistically coherent sentences.

Methodological Overview

The proposed method innovatively segments the problem of text style transfer into distinct phases. Initially, the "Delete" phase involves identifying and removing elements in a sentence that convey the original sentiment, thereby neutralizing the text. The model then enters the "Retrieve" stage, wherein it sources phrases from a corpus that align with the desired sentiment or style. Lastly, the "Generate" phase synthesizes a coherent sentence by integrating the retrieved elements with the neutralized content.

Experimental Results

On benchmark datasets, this method demonstrated competitive performance in altering sentiment while preserving semantic content. Notably, the approach managed to achieve substantive results with relatively simple mechanisms compared to more complex neural models prevalent in the field. Quantitative evaluations, including BLEU scores and human judgment on sentiment accuracy and content preservation, reflect the efficacy of this streamlined approach.

Theoretical and Practical Implications

The primary contribution of this research lies in its decomposition of style transfer into intuitive, explainable steps, offering an interpretable alternative to end-to-end neural systems. The methodology underscores the potential for modular design in natural language processing tasks, suggesting avenues for enhancing control and transparency in model predictions.

Moreover, from a practical standpoint, this model's reliance on non-specialized components and manageable computational overhead makes it accessible for real-world applications where resources may be constrained.

Future Directions

Future research could explore the integration of more sophisticated retrieval strategies or automated alignment of retrieved phrases with the stylistic goals. Additionally, extending this framework to encompass a wider array of stylistic dimensions beyond sentiment could bolster its applicability. Investigating the dynamics of these modules within more diverse linguistic contexts could further reveal limitations or adaptations necessary for broader language coverage.

In summary, the "Delete, Retrieve, Generate" framework offers a compelling direction for sentiment and style transfer, balancing simplicity with performance. Its contributions to the field have significant implications for both the theoretical understanding and practical execution of text manipulation tasks in natural language processing.