BibSonomy :: publication ::

tag user group author concept BibTeX key search:all search:diego_ma

The blue social bookmark and publication sharing system.

(0)

This publication has not been reviewed yet.

rating distribution

average user rating

The average rating is computed over all reviews. However, some of them may be invisible to you due to the visibility setting chosen by the reviewers.

(0.0 of 5.0 based on 0 reviews)

Automatically Generating Wikipedia Articles: A Structure-Aware Approach

by: Christina Sauper, and Regina Barzilay

In: Proceedings ACL 2009 (2009) .

Resources (URL, PDF, PS...)

URL:	http://people.csail.mit.edu/regina/papers.html

In this paper, we investigate an approach for creating a comprehensive textual overview of a subject composed of information drawn from the Internet. We use the high-level structure of human-authored texts to automatically induce a domain-specific template for the topic structure of a new overview. The algorithmic innovation of our work is a method to learn topic-specific extractors for content selection jointly for the entire template. We augment the standard perceptron algorithm with a global integer linear programming formulation to optimize both local fit of information into each topic and global coherence across the entire overview. The results of our evaluation confirm the benefits of incorporating structural information into the content selection process.

BibTeX record

@inproceedings{Sauper:2009,
  abstract = {In this paper, we investigate an approach for creating a comprehensive textual overview of a subject composed of information drawn from the Internet. We use the high-level structure of human-authored texts to automatically induce a domain-specific template for the topic structure of a new overview. The algorithmic innovation of our work is a method to learn topic-specific extractors for content selection jointly for the entire template. We augment the standard perceptron algorithm with a global integer linear programming formulation to optimize both local fit of information into each topic and global coherence across the entire overview. The results of our evaluation confirm the benefits of incorporating structural information into the content selection process.},
  added-at = {2010-04-30T04:32:30.000+0200},
  author = {Sauper, Christina and Barzilay, Regina},
  biburl = {http://www.bibsonomy.org/bibtex/202274f6855f0778829bcc808f7ffc365/diego_ma},
  booktitle = {Proceedings ACL 2009},
  interhash = {6e32e4cc7ae5056581861cf35cd56a9d},
  intrahash = {02274f6855f0778829bcc808f7ffc365},
  keywords = {summarisationmultidocumentgenerationwikipedia},
  library = {Mine (March 2010)},
  review = {There are two key aspects that are proposed:  1. Template induction: Automatically induce structure (the sections of the summary). 2. Search for text for each component of the structure. The method accounts for interdependence between sections.  1 TEMPLATE INDUCTION  The main idea is to cluster the section headings for all documents. The largest clusters are selected to become topics.  To cluster the section headings they follow a repeated bisectioning algorithm using a TF.IDF-weighted cosine similarity measure. Clusters with low internal similarity (according to a threshold, apparently 0.5) are discarded. The number k of sections chosen is computed as the average of sections in all documents. The sections are ranked using a majority ordering algorithm. Each topic is labeled with the most frequent heading of the cluster. The result is a list of topics (t1, ... tk).  2 FILLING TEMPLATES  For each topic, candidate text is searched by combinging the document title and the topic (e.g. "3-M syndrome" diagnosis). The top 10 hits are selected and segmented into chunks delimited by <p> and similar markup. Note that this search is done on Wikipedia. The result is a list of excerpts (ej1, ... ejr) for each topic tj.  Patterns are selected simultaneously from all topics using a joint training algorithm based on perceptron. The process to learn the model is:   a. Rank each candidate excerpt based on how representative it is of its assigned topic.  b. Select excerpts that keep the global rank of all topics minimal. For this they use Integer Linear Programming (ILP):        - Function to minimise: the sum of the ranks of the selected excerpts.        - Constraints to meet:             * Each topic has at least one excerpt.             * No pair of selected excerpts has a similarity above 0.5 (using cosine similarity). So if to exerpts have a similarity over 0.5, at most one of them will be selected.      The features used in the ILP method are as used in other summarisation methods (see Table 1).  For part b, each topic tj is assigned a weight vector wj by using a modification of the perceptron ranking algorithm. The process assumes that each section of each document has a gold excerpt added to it (I think this is taken from the sample solution).  EVALUATION  The system is evaluated against the following baselines and an oracle:  - Baseline 1: Select the top page from a web search and select the first k paragraphs (same k as above).  - Baseline 2: Select paragraphs according to machine learning trained on positive and negative samples. Choose those k excerpts with low redundancy (using a greedy approach?)  - Baseline 3: As proposed in this paper but without performing the optimisation step; simply select the highest-ranking excerpt for each topic.  - Oracle: Select the excerpts that are closest (cosine similarity) to the sample solution.  There are two evaluations:   - ROUGE.   - Analysis of human edits of the proposed solutions.},
  timestamp = {2010-04-30T04:32:30.000+0200},
  title = {Automatically Generating Wikipedia Articles: A Structure-Aware Approach},
  url = {http://people.csail.mit.edu/regina/papers.html},
  year = 2009
}

Endnote record

%0 Conference Paper
%1 Sauper:2009
%A Sauper, Christina
%A Barzilay, Regina
%B Proceedings ACL 2009
%D 2009
%K 
%T Automatically Generating Wikipedia Articles: A Structure-Aware Approach
%U http://people.csail.mit.edu/regina/papers.html
%X In this paper, we investigate an approach for creating a comprehensive textual overview of a subject composed of information drawn from the Internet. We use the high-level structure of human-authored texts to automatically induce a domain-specific template for the topic structure of a new overview. The algorithmic innovation of our work is a method to learn topic-specific extractors for content selection jointly for the entire template. We augment the standard perceptron algorithm with a global integer linear programming formulation to optimize both local fit of information into each topic and global coherence across the entire overview. The results of our evaluation confirm the benefits of incorporating structural information into the content selection process.

BibSonomy :: publication ::

Automatically Generating Wikipedia Articles: A Structure-Aware Approach

Resources (URL, PDF, PS...)

Abstract

BibTeX record

Endnote record