Difference between revisions of "Bethard cikm2010"

From Cohen Courses
Jump to navigationJump to search
m
m
Line 53: Line 53:
 
== Experiments ==
 
== Experiments ==
 
Feature scores were log-transformed and scaled to between 0 and 1. <br>
 
Feature scores were log-transformed and scaled to between 0 and 1. <br>
Experimented with two classifiers :
+
They experimented with two classifiers :
 
* [[UsesMethod:: Logistic_regression]]
 
* [[UsesMethod:: Logistic_regression]]
 
* [[UsesMethod::Support_Vector_Machines]]-MAP
 
* [[UsesMethod::Support_Vector_Machines]]-MAP

Revision as of 17:53, 1 April 2011

Citation

  • Title : Who Should I Cite? Learning Literature Search Models from Citation Behavior
  • Authors : S. Bethard and D. Jurafsky
  • Venue : CIKM 2010

Summary

This paper describes a retrieval model to search relevant existing work in a collection of scientific articles. The authors claimed that the model is useful when a researcher wants to conduct a new research outside his/her area of expertise and needs to get familiar with prior work in the field. The model incorporates various text and meta features and uses the citation networks to learn weights of these features.

Dataset

ACL Anthology (~11,000 papers)

Model

Documents are ranked based on their scores.
For scoring, they used linear model between a query Q (project idea) and a document D (existing scientific article) as follows :

Features

To compute the score between a query Q and a document D, they used the following features :

  • Terms :
    • TF-IDF between Q and D
  • Citations :
    • Number of papers that cited D
    • Number of citations for articles in the venue in which D was published
    • Number of citations author of D has received (if there are multiple authors, use one with the most citation counts)
      • A variant using h-index instead of raw citation counts was also explored. An author with h-index h has published h papers each of which has been cited at least h times.
    • PageRank score of document collection calculated over the citation network (instead of the hyperlink network)
  • Recency:
    • year Q - year D. (effect : older papers get less scores).
  • Cited using similar terms :
    • TF-IDF between Q and all documents cited D
    • TF-IDF between Q and a vector constructed using PMI to select important terms used when other documents cited D
  • Similar Topics (100 topics from LDA) :
    • Cosine similarity between Q and D
    • Cosine similarity between Q and averaged topic distributions of all other documents which cited D
    • Topic citation count score
      • For each document in the collection, choose its most prominent topic (denoted by T). Documents cited by this document are considered to be cited by topic T. Normalized the number of topic citations for each document to get the probability. For a query Q, choose its most prominent topic, denoted by S. Topic citation score is the probability computed above for document D and topic S.
    • Entropy of D's topic distribution
    • Entropy of documents-citing-D's mean topic distributions
  • Social habits :
    • Authors : boost D if it has been cited by authors of Q
    • Authors-cited-article : boost D if it has been cited by authors of Q
    • Authors-cited-author : boost D if it was written by authors that were cited by authors of Q
    • Authors-cited-venue : boost D if it appeared in the venue authors of Q has cited
    • Authors-coauthored : boost D if it was written by authors who have co-authored with authors of Q.

Experiments

Feature scores were log-transformed and scaled to between 0 and 1.
They experimented with two classifiers :

Training / Dev / Test split
Bjtdtsplit.png

Results

  • Mean average precision on the dev set for different classifiers (using all features)

Logistic 50/50 model downsampled the number of negative examples to be the same as the number of positive examples. Bjdevres.png

  • Mean average precision on the dev and test sets using different feature sets (model : SVM-MAP)

Bjtestres.png

Feature analysis

Below are feature weights from the SVM-MAP model used to evaluate the test set
Bjfeatanalysis.png