Object Recognition Base on Deep Belief Network
Yajun Zhang & Zongtian Liu & Wen Zhou
School of Computer Engineering and Science
Shanghai University
Shanghai, China
e-mail: zyj1985email@163.com
e-mail: ztliu@shu.edu.cn
e-mail: zhouwen@shu.edu.cn
Yalan Zhang
Mechanical and electrical engineer division
Wenhua College
Wuhan, China
yalanz@126.com
Abstract—Event ontology is a general knowledge base constr-
ucted by event as the basic knowledge unit for computer com-
munication. Event contains six elements which are action,
object, time, environment, assertion and language performance.
In this paper, we mainly discuss object elements recognition.
There are several mainly existing way to recognize object:
methods based on rule, statistical and shallow machine
learning. Although these methods can get better recognition
results in a particular environment, they have nature defects.
For instance, it is difficult for them to do feature extraction
and they can not achieve complex function approximation,
leading to low recognition accuracy and scalability. Aiming at
problems of existing object recognition methods, we present a
Chinese emergency object recognition model based on deep
learning (CEORM). Firstly, we use word segmentation system
(LTP) to segment sentence, and classify words according to
annotating elements in CEC2.0 corpus, and then obtain each
word’s vectorization of multiple features, which include part of
speech, dependency grammar, length, location. We obtain
word’s deep semantic characteristics from the collection after
vectorization using deep belief network, finally, object
elements are classified and recognized. Extensive testing
analysis shows that our proposed method can achieve better
recognition effect.
Keywords-event ontology; deep learning; event recognition;
DBN; CEORM
Event is perceived, relatively independent and moving
existence, thus there is essential distinction between event
and static concept [1]. The event development and evolution
process involves many entities defined as event elements,
which contains action, object, time, location et. Event
element recognition is the most basic and important task of
nature language processing system based on event, which
aims at displaying unstructured text as structured text. We
mainly discuss object element recognition in this paper.
Among event related elements recognition, object
recognition is more complex and difficult than other element
recognition due to object element own features. And its main
features are as following: It has complex structure; Its’
appearance frequency is high in a sentence, which increases
complexity of object element recognition. There
are several mainly existing way to recognize object: methods
based on rule, statistical and shallow machine learning,
which have nature defects, for instance, they can not realize
complex function approximation and cause derivative
withering away, leading to low accuracy of recognition.
In recent years, deep learning (DL) appeared in machine
learning field, which can realize complex function
approximation by learning a deeper nonlinear network
structure, and represent distributed data and obtain feature
function of high-dimensional data [2]. Building many hidden
layers of machine learning model and massive training data,
learning more abstract features, ultimately the accuracy of
classification or prediction will be improved. Aiming at
above problems, we propose a new object recognition
method based on DL. After researching deep learning
mechanism, we explore semantic feature generalization
expression for object-oriented recognition, and then mine
deep semantic information and research its role plays in
object recognition. Using feature analysis method, this paper
abstracts feature from four aspects for word in text;
Analyzing Chinese emergency object recognition model
based on deep learning (CEORM), recognition results are
obtained. Comparative experimental analysis shows that this
method is more effective than traditional shallow network
method.
Prevalent event element recognition methods are mainly
divided into two types: methods based on models, and those
based on machine learning. Model based models is primarily
manually designed custom event models that use a variety of
pattern matching algorithms to match the text with the
custom model. For example, reference paper [3] proposed
football event information extraction system, and reference
paper [4] developed a meteorology event extraction system
based on ontology. Methods based on machine learning
focus on the construction of a classifier and the discovery of
characteristics, combination, and selection. These methods
approach event argument recognition as a classification
problem and select appropriate features for their
classification. In 2002, Hai firstly introduced a maximum
entropy classifier [5] for event extraction and used it to
recognize event arguments.