The Pragmatics Information Extraction
Based on BP Neural Network
Liu Ding, Jiang Minghu
School of Humanities and Social Sciences, Lab. of Computational Linguistics,
Tsinghua University, Beijing , China
liuding10@mails.tsinghua.edu.cn, jiang.mh@tsinghua.edu.cn
Abstract—This article describes a method that uses the BP
neural network to extract the pragmatics information from the
conversational corpus. Then cluster the texts based on the
pragmatics information matrix that generated by the BP neural
network. And, compare with the original term-document matrix,
the pragmatics information matrix improve the clustering results
obviously.
Key words: BP neural network, pragmatics, conversational
corpus, text clustering
I. INTRODUCT ION
In the recent study on nature language processing, the
researcher use lexical information, syntax information and
sematic information, rather than pragmatics information. But
in some research areas, people have to deal with the high
dimensional and sparse matrix, especially in text classification
and text clustering. The high dimensional and sparse matrix
not only contains a great many redundant information, but also
processed by more computer resources. Actually, people do
not need the huge high dimensional and sparse matrix so called
term-document matrix when cluster or classify the texts.
Indeed, there are some special key words which are enough to
identify the feature of the sample in texts, such as the time
words, scene words and role words in conversational texts.
These kinds of information are related to the participator in
conversation, therefore it could be viewed as the pragmatics
information, and this article describes the method that extracts
the pragmatics information.
II. P
RAGMAT ICS INFORMAT ION EXT RACT IONS
The direct method that extracts the time, scene and role
information from conversational texts is the name-entity
recognition technology. But it does not run well because of the
complexity of the oral context in conversational texts. And, the
pragmatics information is often implied by the context, rather
than showed directly by the words. At this time, the
name-entity recognition technology fails. Therefore, we use
the supervised machine learning model -- Back Propagation
Neural Network (BP network) to extract the pragmatics
information from conversational text. The BP network is a
kind of classical learning model, and it is often applied in
pattern recognition. Backpropagation is the generalization of
the Widrow-Hoff learning rule to multiple-layer networks and
nonlinear differentiable transfer functions. Input vectors and
the corresponding target vectors are used to train a network
until it can approximate a function, associate input vectors
with specific output vectors, or classify input vectors in an
appropriate way as defined by user. Standard backpropagation
is a gradient descent algorithm, as is the Widrow-Hoff learning
rule, in which the network weights are moved along the
negative of the gradient of the performance function[12]. The
basic structure of BP neural network is shown as follows:
Input
layer
Hidden
layer
Output
layer
Figure 1. BP Neural Network
Multilayer networks often use the log-sigmoid function
(logsig) and the tan-sigmoid function (tansig) as the transfer
function. The diagrams of these two transfer function are
___________________________________
978-1-4673-2197-6/12/$31.00 ©2012 IEEE