Conditional Random Fields for Sequential Labelling

Documentation | Code | Applications | Related Work

Introduction

We aim at developing CRF tools for sequential labelling. Models to be implemented:

• Standard first, second-order CRFs.

• Semi-Markov CRFs.

• Standard first, second-order MEMM.

Documentation

• A Tutorial on the Maths behind Conditional Random Fields for Sequential Labelling [pdf]
• A Practitioner Guide to Conditional Random Fields for Sequential Labelling [pdf]

Code

• CRF-SL/C++ version 1.0.1  (18/1/2010). Main features:
  • A light weight, generic C++ implementation. Domain independent.
  • First & second-order sequential CRFs.
  • Training algorithms: L-BFGS (from Taku Kudo), stochastic gradient ascent and Collin's voted perceptron.
  • Handle missing training labels.
  • Decoding: Viterbi (forward/backward) and Pearl's max-product.
  • Report label-wise and segment-wise statistics.

Applications

• English noun-phrase chunking: pre-computed data features for the CoNLL-2000 shared task.

Related Work

• FlexCRF package written in C/C++ by Xuan-Hieu Phan, Le-Minh Nguyen and Cam-Tu Nguyen.
• A Toolbox written in Matlab/C by Kevin Murphy. Supports generic CRFs (not limited to sequential structures).
• CRF++ a toolkit , written in C/C++ by Taku Kudo.
• A CRF package written in Java by Sunita Sarawagi. Supports standard CRFs and Semi-Markov CRFs.
• MALLET package written in Java, a development lead by Andrew McCallum. This is a generic machine learning tool for textual data.
• Stanford Named Entity Recognizer with CRF written in Java by Jenny Finkel.

Last Update: 18/1/2010