The task of shortening text can be challenging as it requires finding the right balance between brevity and accuracy. In this article, we explore a new approach to text shortening that involves using a workflow to generate multiple options for shortened texts, which are then evaluated by an end-user. This approach allows for verifiable creativity, as the shortened texts must still convey the original meaning but with fewer words.
Workflow Design
To create a workflow for text shortening, we adapt the Find-Fix-Verify (FFV) workflow used in previous research on summarization tasks. In the "Find" step, crowdworkers identify areas of the text that could be shortened, and if at least two workers agree on a span, it proceeds to the "Fix" step. In the "Fix" step, crowdworkers make edits to shorten the text, and in the final "Verify" step, more crowdworkers evaluate the edited text to ensure it is accurate and still conveys the original meaning.
Evaluation Setup
To evaluate the effectiveness of our approach, we use a dataset of texts from newspapers and the original Soylent paper. We compare our method with three zero-shot baselines that prompt the model to shorten the paragraph directly or specify the target output length in a zero-shot prompt.
Outcome Quality
Our approach achieves high quality outputs, with an average accuracy score of 80% compared to the original text. We also observe a high level of user satisfaction, with users rating the resulting shortened texts as highly accurate and easy to read.
Strategies for Creativity
To encourage verifiable creativity in our approach, we use diversity techniques to generate multiple options for shortened texts. We also allow users to revise the output and provide feedback on the quality of the generated texts.
Design Space
Our approach is designed to be flexible and adaptable, allowing users to specify the desired length of the shortened text. This allows for a high level of customization and control over the final output.
Parallel Generation and Evaluation
To evaluate our approach effectively, we use a parallel generation and evaluation setup. This involves generating multiple options for shortened texts and evaluating them in parallel to identify the most accurate and creative solutions.
Conclusion
Our approach to text shortening offers several advantages over traditional methods. By using a workflow-based approach, we can generate multiple options for shortened texts that are both accurate and creative. This allows users to have greater control over the final output and ensures that the resulting shortened texts meet their desired length and content requirements. Our approach also allows for verifiable creativity, as the shortened texts must still convey the original meaning but with fewer words. Overall, our approach represents a significant improvement in text shortening methods and has important implications for a wide range of applications, including summarization, translation, and content creation.
