Invention Title:

Revision of and Attribution for Output of Text Generation Models

Publication number:

US20260339990

Publication date:
Section:

Physics

Class:

G06F16/33295

Inventors:

Applicant:

Smart overview of the Invention

The patent application addresses the limitations of existing language models (LMs) in generating reliable content. It introduces systems and methods to enhance the accuracy and attribution of LM-generated text. By leveraging machine-learned models, the system can edit content and attribute parts of it to evidence snippets, improving factual accuracy and allowing users to verify claims against external evidence.

Methodology

The proposed method involves obtaining initial textual content from a language model and then retrieving relevant evidence sources. A comparison is made between the generated content and these evidence sources using another language model. This process results in a revised text that is more aligned with the evidence. Portions of the evidence are linked as attributions to the generated content, ensuring transparency and reliability.

Evidence Retrieval

Retrieving evidence involves generating queries based on the initial content. These queries are used to perform searches, potentially across the web, to gather relevant documents. A third language model may assist in formulating these queries by identifying claims within the content. The evidence snippets are then extracted, scored for relevance, and selected to maximize coverage of the claims.

Content Comparison and Revision

The comparison process assesses the agreement between the generated content and evidence snippets. A machine-learned agreement model evaluates this alignment, potentially employing techniques like chain-of-thought or few-shot prompting. If discrepancies are found, a second language model generates revised content with improved agreement to the evidence.

Output Evaluation

The system determines whether to output the revised content based on metrics such as edit distance, agreement score, and preservation score. These metrics ensure that the revised content not only aligns better with the evidence but also retains the essence of the original text. This process results in more reliable and verifiable LM-generated content.