Discriminative Models for Information Retrieval

PubDate(2004), PubPlace(SIGIR) Author(Nallapati)
keyword(Discriminative Model,SVM,Maximum Entropy)

Summary

Applying discriminative model to IR task resulted in improved result for Homepage finding task that typically requires combining many arbitrary features.

Content

Background

• People most have approached IR using generative classifier
  • LMIR : Assuming that each document forms its own class, it’s similar to classifying given query to best-matching document class using Naive-bayes classifier

Contribution

• Problem for IR as classification
  • Out-of-vocabulary problem : query-word may not appear in document
    • Overcome by aggregate feature
  • Unbalance of relevance/non-relevance class
    • Overcome by undersampling nr class

Experiment

• Application of SVM to

Future Work

Comment

Reference

Tags : Paper,IR Print Comments Trackback