# Retrieving Texts based on Abstract Descriptions

**Authors:** Shauli Ravfogel, Valentina Pyatkin, Amir DN Cohen, Avshalom Manevich, Yoav Goldberg  
**arXiv:** [2305.12517](https://arxiv.org/abs/2305.12517)  
**Date:** 2023-05-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.12517v1/

## Key points

- The preprint explores the task of retrieving sentences from large text collections based on abstract descriptions or specifications.
- Existing retrieval-based models and instruction-tuned Large Language Models (LLMs) have limitations in performing semantic retrieval.
- LLMs are not suitable for similarity search over embedding vectors, while keyword-based retrieval methods rely on exact lexical matches which makes them weak for retrieval based on abstract descriptions.
- The authors propose a description-based retrieval approach that leverages the strengths of LLMs to generate descriptions of sentences sampled from Wikipedia.
- They produce five different descriptions for each text, in addition to incorrect descriptions, to be used as negative examples.
- The resulting dataset includes both positive and misleading descriptions that align with the original sentence and the abstract description.
- The authors train a descriptions encoder and a text encoder using positive and negative pairs sourced through prompting an LLM.
- The vector encodings learned by these encoders can be used in a standard similarity-based retrieval setting to retrieve sentences that align with a user's description or specification.
- The proposed model significantly improves upon existing text embeddings when used in standard nearest neighbor search.
- This approach can serve as a useful component to enhance discovery ability in many data-intensive domains, especially in professional domains such as legal, medical or scientific search.

## Summary

This preprint explores the task of retrieving sentences from large text collections based on abstract descriptions or specifications. The authors identify the limitations of existing retrieval-based models and instruction-tuned Large Language Models (LLMs) in performing semantic retrieval. While LLMs excel at extracting information from text, they are not suitable for similarity search over embedding vectors. On the other hand, keyword-based retrieval methods rely on exact lexical matches, which makes them inherently weak for retrieval based on abstract descriptions. To address this challenge, the authors propose a description-based retrieval approach that leverages the strengths of LLMs to generate descriptions of sentences sampled from Wikipedia. They produce five different descriptions for each text, in addition to incorrect descriptions, to be used as negative examples. The resulting dataset includes a wide range of both positive and misleading descriptions that align with the original sentence and the abstract description. The authors then train a descriptions encoder and a text encoder using positive and negative pairs sourced through prompting an LLM. The vector encodings learned by these encoders can be used in a standard similarity-based retrieval setting to retrieve sentences that align with a user's description or specification. The proposed model significantly improves upon existing text embeddings when used in standard nearest neighbor search. The results demonstrate that data from LLMs can be used not only for distilling more efficient specialized models than the original LLM but also for creating new capabilities not immediately possible using the original model. The description-based retrieval capability demonstrated in this work can serve as a useful component to enhance discovery ability in many data-intensive domains, especially in professional domains such as legal, medical or scientific search. Overall, this preprint presents an innovative approach to improving semantic retrieval using large language models and highlights its potential applications across various domains.