Seminario MAVIR

Lluís Màrquez: Statistical Learning for Structured Natural Language Processing

12-13 de diciembre de 2006

Lugar de celebración

Resumen

In this seminar, we will present the statistical learning approach for dealing with the annotation of natural language structures. In doing that, we will overview several techniques and methods from the Machine Learning (ML) literature, including the following topics:

• Generative Models

• Discriminative learning for sequential labeling tasks

• The learning-and-inference paradigm (an ILP approach to reason with constraints)

• Re-ranking techniques for working with complete solution candidates

• Global linear models for structure learning, including some SVM-like algorithms to train them

Some problems that might serve as basic examples of the preceding topics are: partial and full parsing, extraction of named entities and relations, and semantic role labeling.

Videos

[googlevideo:http://video.google.es/videoplay?docid=-2854488951646797101&hl=es]

[googlevideo:http://video.google.es/videoplay?docid=-1867269727798040819&hl=es]

[googlevideo:http://video.google.es/videoplay?docid=7599312341689712047&hl=es] 

Materiales

• Presentación de Lluís Màrquez (UPC) Statistical Learning for Structured Natural Language Processing

• Presentación de Michael Collins (MIT) An SVM Approach for Natural Language Learning

• Materiales del seminario práctico (70 MB): software, modelos de entrenamiento, datos, scripts de extracción de rasgos, logs de usuarios, etc.

Software

Las herramientas que aparecen en el paquete de materiales han sido compiladas para i386 en máquinas con procesador Pentium Dual Core. Para hacerlo funcionar en cualquier otro ordenador, es probable que sea necesario recompilar. Las fuentes y las instrucciones para reinstalar están disponibles en:

• TnT
• Yamcha
• svm-hmm
• Mallet
• CRF++