# Extraction of Salient Sentences from Labelled Documents

**Authors:** Misha Denil, Alban Demiraj, Nando de Freitas  
**arXiv:** [1412.6815](https://arxiv.org/abs/1412.6815)  
**Date:** 2014-12-21  
**Summary page:** https://www.summarizepaper.com/en/paper/1412.6815v2/

## Key points

- The paper presents a hierarchical convolutional document model for introspection of document structure
- Visualisation techniques from computer vision are applied to identify and extract topic-relevant sentences using this model
- A new scalable evaluation technique for automatic sentence extraction systems is introduced, eliminating the need for manual annotation of validation data
- The proposed architecture allows for a deeper understanding of the document structure
- The approach provides an efficient alternative for evaluating the performance of automatic sentence extraction systems
- The paper offers valuable insights into extracting salient sentences from labelled documents
- The hierarchical convolutional document model and visualization techniques provide a robust framework for identifying topic-relevant information.

## Summary

The paper titled "Extraction of Salient Sentences from Labelled Documents" presents a hierarchical convolutional document model that supports introspection of the document structure. The authors demonstrate how visualisation techniques from computer vision can be applied to identify and extract topic-relevant sentences using this model. Additionally, they introduce a new scalable evaluation technique for automatic sentence extraction systems, eliminating the need for time-consuming human annotation of validation data. The paper's authors, Misha Denil, Alban Demiraj, and Nando de Freitas, propose a novel architecture that allows for a deeper understanding of the document structure. By leveraging techniques from computer vision literature, they are able to visualize and extract sentences that are most relevant to the given topic. One significant contribution of this work is the introduction of a scalable evaluation technique for automatic sentence extraction systems. This technique eliminates the need for manual annotation of validation data, which can be time-consuming and costly. The authors' approach provides an efficient alternative for evaluating the performance of such systems. Overall, this paper offers valuable insights into extracting salient sentences from labelled documents. The proposed hierarchical convolutional document model and visualization techniques provide a robust framework for identifying topic-relevant information. The introduced evaluation technique enhances scalability and efficiency in assessing automatic sentence extraction systems.