A sparse L2-regularized support vector machines for large-scale natural language learning

Yu Chieh Wu, Yue Shi Lee, Jie Chi Yang, Show Jane Yen

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Linear support vector machines (SVMs) have become one of the most prominent classification algorithms for many natural language learning problems such as sequential labeling tasks. Even though the L2-regularized SVMs yields slightly more superior accuracy than L1-SVM, it produces too much near but non zero feature weights. In this paper, we present a cutting-weight algorithm to guide the optimization process of L 2-SVM into sparse solution. To verify the proposed method, we conduct the experiments with three well-known sequential labeling tasks and one dependency parsing task. The result shows that our method achieved at least 400% feature parameter reduction rates in comparison to the original L2-SVM, with almost no change in accuracy and training times. In terms of run time efficiency, our method is faster than the original L2-regularized SVMs at least 20% in all tasks.

Original languageEnglish
Title of host publicationInformation Retrieval Technology - 6th Asia Information Retrieval Societies Conference, AIRS 2010, Proceedings
Pages340-349
Number of pages10
DOIs
StatePublished - 2010
Event6th Asia Information Retrieval Societies Conference, AIRS 2010 - Taipei, Taiwan
Duration: 1 Dec 20103 Dec 2010

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume6458 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference6th Asia Information Retrieval Societies Conference, AIRS 2010
Country/TerritoryTaiwan
CityTaipei
Period1/12/103/12/10

Keywords

  • L-regularization
  • machine learning
  • part-of-speech tagging
  • support vector machines

Fingerprint

Dive into the research topics of 'A sparse L2-regularized support vector machines for large-scale natural language learning'. Together they form a unique fingerprint.

Cite this