Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TOP:A New Target-Audience Oriented Content Paraphrase Task

Boda Lin, Jiaxin Shi, Haolong Yan, Binghao Tang, Xiaocheng Gong, Si Li

TL;DR

This work introduces TOP, a content paraphrase task designed to tailor outputs to a target audience using user history and generation-capable models. It formalizes a unified framework with preference extraction/encoding and content generation, and provides the Customized-TOP multimodal dataset to support text and image paraphrasing. The authors implement baseline TOP pipelines for text and image modalities and propose three evaluation metrics—Content Preservation Index, Preference Preservation Index, and Natural Realism Index—to capture different facets of quality. Results indicate promising content preservation and realism for text, and strong style-transfer performance for images, while highlighting areas for improving preference alignment and concept transfer in vision tasks. These contributions offer a pathway to more personalized and context-aware recommendation experiences.

Abstract

Recommendation systems usually recommend the existing contents to different users. However, in comparison to static recommendation methods, a recommendation logic that dynamically adjusts based on user interest preferences may potentially attract a larger user base. Thus, we consider paraphrasing existing content based on the interests of the users to modify the content to better align with the preferences of users. In this paper, we propose a new task named Target-Audience Oriented Content Paraphrase aims to generate more customized contents for the target audience. We introduce the task definition and the corresponding framework for the proposed task and the creation of the corresponding datasets. We utilize the Large Language Models (LLMs) and Large Vision Models (LVMs) to accomplish the base implementation of the TOP framework and provide the referential baseline results for the proposed task.

TOP:A New Target-Audience Oriented Content Paraphrase Task

TL;DR

This work introduces TOP, a content paraphrase task designed to tailor outputs to a target audience using user history and generation-capable models. It formalizes a unified framework with preference extraction/encoding and content generation, and provides the Customized-TOP multimodal dataset to support text and image paraphrasing. The authors implement baseline TOP pipelines for text and image modalities and propose three evaluation metrics—Content Preservation Index, Preference Preservation Index, and Natural Realism Index—to capture different facets of quality. Results indicate promising content preservation and realism for text, and strong style-transfer performance for images, while highlighting areas for improving preference alignment and concept transfer in vision tasks. These contributions offer a pathway to more personalized and context-aware recommendation experiences.

Abstract

Recommendation systems usually recommend the existing contents to different users. However, in comparison to static recommendation methods, a recommendation logic that dynamically adjusts based on user interest preferences may potentially attract a larger user base. Thus, we consider paraphrasing existing content based on the interests of the users to modify the content to better align with the preferences of users. In this paper, we propose a new task named Target-Audience Oriented Content Paraphrase aims to generate more customized contents for the target audience. We introduce the task definition and the corresponding framework for the proposed task and the creation of the corresponding datasets. We utilize the Large Language Models (LLMs) and Large Vision Models (LVMs) to accomplish the base implementation of the TOP framework and provide the referential baseline results for the proposed task.
Paper Structure (20 sections, 14 equations, 6 figures, 3 tables)

This paper contains 20 sections, 14 equations, 6 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLMs
  4. Diffusion Models
  5. Recommendation System
  6. Preliminaries
  7. Task Definition
  8. Unified Framework
  9. Metric Design
  10. Implementation
  11. TOP on Text
  12. TOP on Image
  13. Datasets
  14. Text Datasets
  15. Image Datasets
  16. ...and 5 more sections

Figures (6)

  • Figure 1: Compared to the approach of directly recommending content to users, the proposed TOP task places a greater emphasis on appropriately paraphrasing existing content based on the preferences of users.
  • Figure 2: The total framework of the proposed TOP task. The user preference is extracted by the preference extractor and encoded by the preference encoder. And the final customized output is generated by the content generator.
  • Figure 3: The figure of our base implementation of the TOP framework. Panel (a) illustrates the overall framework, which aligns with the TOP framework described in Section 3. Panel (b) delineates the specific implementation of TOP for text. Historical data is concatenated with a preference prompt and then fed into a preference LLM to extract the preferences. The content text is concatenated with a paraphrase prompt and in conjunction with the extracted preferences is processed through a paraphrase LLM to generate the outcome. Panel (c) demonstrates the specific implementation of TOP for images. (c-1) depicts the implementation of TOP applied to the task of image style transfer and (c-2) depicts the implementation of TOP applied to the task of iamge concept transfer.
  • Figure 4: Display of the original source of text data.
  • Figure 5: Teasers of TOP on Image for Style Transfer. It can be observed that due to the histogram loss and the reduction of the weight of the style loss, our generation results do not produce texture distortion which is common in photo style transfer.
  • ...and 1 more figures