Browse Prior Art Database

A Maximum-Entropy Based Confidence Metric for a Statistical Parser

IP.com Disclosure Number: IPCOM000028875D
Original Publication Date: 2004-Jun-05
Included in the Prior Art Database: 2004-Jun-05
Document File: 1 page(s) / 22K

Publishing Venue

IBM

Abstract

This is a method used to apply maximum entropy methods and feature selection similar to that proposed in [2] to parse trees built using the algorithm of [1]. This yields a good confidence metric that can also be used to select between the top-N candidates that the method of [1] produces.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

Page 1 of 1

A Maximum-Entropy Based Confidence Metric for a Statistical Parser

A conversational system usually performs the 'understanding' of what the user said using a 2 step approach. First a natural language understanding component extracts meaning from a sentence, typically represented in predicate calculus or attribute-value notation. Then a dialog manager takes this meaning and the context of the conversation, and produces a response for the user. Many approaches have been used for natural language understanding, a common one being statistical parsing [1], [2]. In a statistical parser, a parse tree is built bottom up, using conditional probability models to guide the creation of constituents from the words. The parse tree is then traversed to produce the predicates or attribute-value list that the dialog manager uses. One challenge with this approach is recognizing when the bottom-up statistical parser has made a mistake. This can happen for many reasons, and depends on the modeling technique used to guide the conditional probability models governing the creation of extensions. If the system recognizes that a parse tree is incorrect, then perhaps the 2nd or 3rd choice parse tree might actually be correct for the sentence. This is a method used to apply maximum entropy methods and feature selection similar to that proposed in [2] to parse trees built using the algorithm of [1]. This yields a good confidence metric that can also be used to select between the top-N candidates that the method of
[1] produces.

Maximum entropy features are selected by applying the phrases dominated by labels in the parse tree along with the attributes that these labels generate. Two types of features are used: exact phrase features, and bag of semantic word features. The exact phrase feature consis...