An Event Element Extraction Method for Chinese Text

Xia Jing

, Wei Zhang

, Jingjing Wang

and Yongli Wang

School of Information Engineering, Nanjing Audit University, NAU Nanjing, Jiangsu 211815, China

School of Computer Science and Engineering, Nanjing University of Science and Technology,

NJUST, Nanjing, Jiangsu 210094, China

Keywords: Event Extraction, Deep Learning, Event Element Detection.

Abstract: With the rapid development of computer technology and Internet scale, how to extract useful information

from the growing mass of network information and present it in the form of structured text is particularly

important. As a solution to this problem, information extraction technology has attracted much attention.

Among them, event extraction is an important research direction in the field of information extraction, and it

is also one of the most challenging tasks. There are some problems in traditional event extraction methods,

such as easy to ignore the context information and insufficient extraction of key features. In order to solve the

above problems, this paper uses deep learning method to study the event extraction of Chinese text, and

proposes the recognition and classification of Chinese event elements, that is, the detection of Chinese event

elements. Using the type information and the corresponding location information of event trigger words, the

text vector is obtained as the input of BiLSTM network layer. The attention layer is added on the basis of

BiLSTM network layer to better obtain the information of event elements around the trigger words. Finally,

the detection results of event elements are obtained through softmax layer output.

1 INTRODUCTION

1.1 Research Background and

Significance

With the development and progress of computer

technology and the scale of the Internet, the

transmission of information has become convenient

and rapid. A large amount of data is produced in the

network every day. People need to face and deal with

complex information every day, which is beyond the

scope of manual processing. Information extraction

technology has been widely studied by scholars and

made rapid progress. How to quickly and accurately

locate the information that users are interested in and

concerned about from the huge network information

flow has become one of the important information

extraction tasks in the field of natural language

processing (Guo, He, 2015).

Among them, event extraction is an important

research direction in information extraction task, and

it is also one of the most challenging tasks in the field

of information extraction. It provides theoretical and

technical support for intelligent question answering,

information search, automatic summarization and

knowledge mapping. At the same time, it is closely

related to data mining, machine learning and other

fields, and promotes the development of related

disciplines (Su, 2017).

Through the work of event extraction, we can get

a complete description of the event, including the

time, place and people involved in the event. Event

extraction technology is also widely used in finance,

medical, judicial and other industries, which provides

a convenient and fast tool for the staff of these

industries, and also improves the corresponding work

efficiency. At present, most of the researches on event

extraction task at home and abroad are oriented to

English text, and some research progress has been

made. Due to the high complexity of Chinese

language, Chinese corpus is relatively small, and the

research on event extraction for Chinese text is just

beginning. Therefore, further research on Chinese text

event extraction is challenging and significant.

1.2 Research Status at Home and

Abroad

With the development of information extraction

technology, event extraction technology has become a

research hotspot and difficulty in the field of natural

Jing, X., Zhang, W., Wang, J. and Wang, Y.

An Event Element Extraction Method for Chinese Text.

DOI: 10.5220/0011753300003607

In Proceedings of the 1st International Conference on Public Management, Digital Economy and Internet Technology (ICPDI 2022), pages 619-626

ISBN: 978-989-758-620-0

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

619

language processing. Scholars at home and abroad are

the first to study event extraction in English text. With

the development of related theory and technology, the

research of English event extraction has achieved

corresponding results, and some scholars begin to

study event extraction in Chinese text. According to

the research progress of relevant scholars in recent

years, the main related methods of event extraction

can be divided into three types: pattern matching

based method, machine learning based method and

deep learning neural network based method. Among

them, the method based on pattern matching can

achieve satisfactory event extraction results in specific

fields, but the method has poor portability and needs

domain experts to make rules. With the improvement

of computer hardware level, event extraction methods

based on machine learning and deep learning have

become the mainstream research direction.

1.3 Main Research Contents of this

Paper

In this paper, Chinese event element detection is

studied. Firstly, the task definition of event element

detection is introduced. Then, the text vector

representation with event trigger word information is

introduced. Then, the Chinese event element detection

model proposed in this paper is introduced in detail.

The attention mechanism is added to the event

detection model, and the information of event trigger

words is used to enhance the detection results of the

model. The rationality and effectiveness of the model

are verified by comparing the proposed method with

the traditional method on the recognized data set.

The main content of this paper is the subtask of

Chinese event extraction: event element extraction. At

present, scholars have done more research on event

trigger words, but the research on event element

detection is limited. Many systems use the same

model for event trigger word detection and event

element detection, and do not redesign the model to

obtain more precise text features.

Based on the combination of BERT model and

recurrent neural network event detection model, this

paper makes adjustments to further improve the

accuracy of event element detection, mainly by adding

the type and location information of trigger words in

the text vector representation of input layer, and

adding attention mechanism in the computing layer of

Bidirectional Long Short-Term Memory (BiLSTM)

network. The specific event element detection model

structure will be described in detail below.

2 MATERIALS AND METHOD

2.1 Event Extraction Definitions

In the field of event extraction, Automatic Content

Extraction (ACE) is the most authoritative

international conference organized by National

Institute of Standards and Technology (NIST) since

2000 (Zhang, 2017). ACE conference defines an event

as an event or a state change that occurs in a specific

time or time range, a specific place or geographical

range, and is composed of one or more participants,

one or more actions (Doddington, Mitchell,

Przybocki, et al., 2004). ACE conference divides the

event extraction task into two sub tasks: the first sub

task is the recognition and classification of events. The

goal of this task is to detect event trigger words from

text data sets and identify their corresponding event

types. The second sub task is to identify and classify

event elements, including time element, place element

and object element. Through the description of the

above related tasks, event extraction is to identify and

classify event information from unstructured or semi-

structured text, and then present it in a structured form

to provide more accurate data for upstream

applications. The concepts related to event extraction

are introduced as follows.

2.1.1 Event Description

The definition of event description refers to the natural

text that describes one or more things, which can be

phrases or sentences, which will contain at least one

event trigger word and one or more event elements.

For the same thing can have different descriptions,

distributed in different texts.

2.1.2 Event Trigger Words

Generally, nouns or verbs are used as event trigger

words, which are the key words to describe an event

and determine the event type. Event trigger word

detection is the first subtask in the event extraction

task.

2.1.3 Event Type

Event type refers to the category of the event itself.

Generally, there are clear definitions of event types in

corpus, such as emergency, mobile event, operation

event and so on. The event type is generally

determined by the event trigger words. The type of

event trigger words is the event type, which is

ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology

620

completed at the same time as the event trigger word

detection.

2.1.4 Event Element

Event element refers to the specific description related

to an event. The specific category can be divided into

time element, location element and object element.

Usually, one or more event elements will be included

in a complete event description. Event element

detection is the second sub task in event extraction

task, which includes the identification and

classification of event elements.

2.2 Introduction to Chinese Emergency

Corpus

Chinese Emergency Corpus (CEC) is an event corpus

for Chinese text. It mainly collects five types of news

reports on the Internet, including earthquake, fire,

traffic accident, terrorist attack and food poisoning, as

the original corpus. The corpus is constructed by the

semantic Intelligence Laboratory of Shanghai

University. Then, we annotate the event, including

event trigger words and event elements. After

consistency checking, the annotated results are saved

and the complete annotated corpus is obtained.

Although the number of texts in this corpus is not very

large, it has the most comprehensive annotation of

event trigger words and event elements, which is very

suitable for Chinese event extraction. Therefore, this

paper uses CEC Corpus as the training set and

verification set of related algorithms.

The latest statistical data of CEC Corpus is shown

in Fig. 1, in which there are 332 texts, 5954 event

trigger words and 14401 event elements. The corpus

uses XML language as annotation format. The main

tags include event, receiver, participant, time and

location.

2.3 Event Element Detection Task

Definition

According to the introduction of events in the previous

Section A, if there is an event e, it can be specifically

defined as e = (A, O, T, V, P, L), where A, O, T, V, P

and L represent action element, object element, time

element, place element, assertion description and

language representation respectively. Action element

is one of the important symbols of event description,

which shows the dynamic of event occurrence. Object

element refers to the people or things involved in the

process of event description, which can be divided

into subject role and object role. Time element refers

to the specific time point of event occurrence or the

time interval of event continuous occurrence. Place

element refers to the location information involved in

the event, including the specific location, such as

Hawaii Island, and the abstract location, such as web

forums. Assertion description refers to the process of

event change, which can be divided into pre assertion,

intermediate assertion and post assertion. Pre assertion

generally refers to the constraint conditions or event

trigger conditions, intermediate assertion generally

refers to the conditions met by each event element in

the process of event, and post assertion refers to the

post condition of event, that is, the change of event

element state after event. Language representation

refers to the linguistic rules that describe events, such

as the common collocations of trigger words. Taking

the sentence "an earthquake occurred in the sea area

near the South Pacific island country of Fiji on the

afternoon of 26 local time" as an example, the

corresponding event element analysis results are

Figure 1: Statistical results of CEC Corpus.

An Event Element Extraction Method for Chinese Text

621

shown in Table 1.

For an event sentence with definite event type, it is

assumed that the event element vocabulary contained

in the sentence is E={e

, e

, …, e

}, where e

is the

i-th event element in the event sentence. Define the

event element category as T={t

, t

, …, t

}, where

represents the i-th category in the event element

category. After the event element detection, the

corresponding event category mapping pair 〈e

, t

〉 can

be obtained, that is, the corresponding type of event

element e

is t

, and the corresponding event category

of e

is unique. The goal of event element detection is

to establish a one-to-one relationship between event

element vocabulary e

and event element type t

This paper focuses on the extraction of time

element, place element and object element in the event

sentence, that is, three element tags corresponding to

time, location and participant in CEC Corpus. For

details, see the introduction of corresponding CEC

Corpus in Section B. Taking "An earthquake occurred

in the sea area near the South Pacific island country of

Fiji on the afternoon of 26 local time" as an example,

after the event detection model processing, the

corresponding event trigger word is "earthquake", and

the event type is "emergency". The task of this paper

is to effectively use the information of trigger words

to detect the event elements contained in sentences. In

this case, it is 〈Fiji, the South Pacific island country,

"Location"〉, 〈" the afternoon of 26", "Time"〉. The

former is the location element, and the latter is the time

element.

Table 1: Events and their Corresponding Event Elements.

Type Corresponding Description

event sentence

an earthquake occurred in the sea

area near the South Pacific island

country of Fiji on the afternoon of

26 local time

action element occurred

object element Fiji, earthquake

time element 26, afternoon

location element the Pacific. Fiji

assertion

descri

tion

post condition: an earthquake

occurre

Language

ression

occurred + earthquake

2.4 Preprocessing of Text Vectorization

Event element detection, like trigger word detection,

is also a sentence level natural language processing

task. Suppose the input sentence is {w

, w

, …, w

where w

is the i-th word in the sentence, and N is the

total number of words in the sentence after word

segmentation. When the sentence is calculated as a

numerical value, the first step is to convert it into the

corresponding text vector, that is, to convert the

sentence into the corresponding vector form {v

, v

…, v

}, where v

is w

is the corresponding word

vector.

Event element extraction is carried out after event

detection, which can make full use of trigger word

information to identify and classify event elements. In

the event element detection model, four kinds of

vectors, namely, the pre training word vector V

, the

part of speech vector V

, the trigger word type vector

and the trigger word position vector V

, are spliced

as the input of the BiLSTM layer, that is, there is (1).

𝑉



= 𝑉



+ 𝑉



+ 𝑉



+ 𝑉



. (1)

Where V

, V

represents the pre training

vector, part of speech vector, trigger word type vector

and trigger word position vector of the BERT model

corresponding to the i-th word, V

is the word vector

corresponding to the i-th word. The following

describes the trigger word type vector and trigger

word position vector.

2.4.1 Trigger Word Type Vector

Through the detailed introduction of CEC Corpus in

Section B, there are eight event types in the corpus,

and nine non trigger word types. Therefore, this paper

uses a 9-dimensional vector to represent the trigger

word type vector, the position of the corresponding

type is set to 1, and other positions are set to 0.

2.4.2 Trigger Word Position Vector

The trigger word position vector records the relative

distance between each word and the trigger word. In

this paper, a 5-dimensional vector is used to represent

the position vector of trigger words, and the maximum

representable distance is 31.

2.5 Chinese Event Element Detection

Model

The combination model of recurrent neural network

and attention mechanism has been successfully

applied to many natural language processing tasks. In

this paper, attention mechanism is introduced into

Chinese event element detection, and the

corresponding changes are made based on the event

detection model combining BERT model and

recurrent neural network. Event element detection is

the second step of event extraction. Using the

ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology

622

information of event trigger words obtained from

event detection task, this paper proposes an event

element detection model combining attention

mechanism and recurrent neural network. The event

detection model is mainly divided into five layers:

input layer, coding layer, attention layer, decoding

layer and output layer.

In this paper, we use the text vector of multi feature

stitching as the input of BiLSTM network layer. After

the first BiLSTM layer, we get the global feature

information of the sentence sequence. Then we input

the calculated results into the attention mechanism

layer to obtain the feature information between words.

Then we input the text information processed by the

attention layer into the second BiLSTM layer. Finally,

the final event element detection result is obtained

through softmax layer.

2.5.1 Input Vector Representation Layer

In the Chinese event element detection model

proposed in this paper, four vectors, namely, the

BERT pre training word vector, the part of speech

vector, the trigger word type vector and the trigger

word position vector, are combined to form a new text

vector. Suppose that the input sentence is S={w

, …, w

}, where w

represents the i-th word in the

sentence after word segmentation, and x

represents

the corresponding word vector of w

, as shown in (2),

where x

, x

respectively represent the pre

training vector, part of speech vector, trigger word

type vector and trigger word position vector of the i-

th word, and x

represents the corresponding word

vector of the i-th word, where x

∈R

. The final

expression of sentence s is shown in (3), where

X∈R

n*d

𝑥



= 𝑥



+ 𝑥



+ 𝑥



+ 𝑥



. (2)

𝑋 =



𝑥



, 𝑥



, 𝑥



,…,𝑥





. (3)

2.5.2 Coding Layer

In the natural language processing task, the coding

layer and decoding layer can choose different

combinations, usually choose the recurrent neural

network as the corresponding information feature

processing layer. In Long Short-Term Memory

(LSTM) model, due to the existence of memory

update unit, the representation of each word combines

the information of the text in front of the word, but this

representation only makes full use of the above

information, and does not make use of the following

information. BiLSTM is proposed to solve the

problem that LSTM can only process text sequence in

one direction. BiLSTM model uses two LSTM

models, which encode the input sequence from the

forward and reverse directions respectively, and then

combine the encoding results of the two directions to

obtain the final representation of each position word.

BiLSTM model has achieved good results in many

natural language processing tasks, especially in tasks

that depend on context global information. The

network structure of BiLSTM is shown in Fig. 2.

As shown in Fig. 2, x

is the word vector input of

the current word, the input layer inputs the vector

representation of each word in the whole sentence, y

represents the final abstract feature representation

after processing by BiLSTM model. For the input

word vector x

. After the forward LSTM network and

the reverse LSTM network processing, the word

vector representation is

ℎ





⃗

and ℎ



⃖



. Then the word

vector input at time t gets the corresponding output

result of context integration, which is shown in (4),

where f represents the text feature information

calculation process of LSTM network layer.

······

−

-1t

−

LSTM LSTM

LSTM

LSTM LSTM LSTM

Figure 2: Schematic diagram of BiLSTM network structure.

An Event Element Extraction Method for Chinese Text

623

𝑦



= 𝑓(𝑥



, ℎ





⃗

, ℎ



⃖



). (4)

2.5.3 Attention Layer

Attention model is to learn the importance of each

input word vector from the sequence, and obtain the

information between words.

In the calculation process of attention mechanism

layer, the (5) is defined, where c

is the context vector

of attention layer, h

is the hidden information ℎ



[ℎ





⃗

, ℎ



⃖



] corresponding to the j-th position in the

BiLSTM network.

The calculation of the corresponding weight of

attention separation a

is shown in (6), and its size

represents the probability of the relationship between

the words in the input sequence and the current output.

Where e

is the attention score, it can be seen that the

value of attention weight a

will increase with the

increase of e

, thus increasing the impact on the final

event element type judgment.

𝑐



∑

𝑎



ℎ







. (5)

𝑎













∑



(





)





. (6)

2.5.4 Decoding Layer

The decoding layer uses BiLSTM network structure,

which is consistent with the network structure of

coding layer. The first hidden state h

of decoding

layer is represented by the last hidden state h

coding layer. The semantic information of sentence

sequence is further extracted by synthesizing the text

feature information processing results of coding layer

and attention layer. The calculation process is shown

in (7), where f represents LSTM network layer.

𝑦



= 𝑓𝑐



, ℎ





⃗

, ℎ



⃖



. (7)

2.5.5 Output Layer

Through the information extraction of BiLSTM

decoding layer, the feature vector H

of the whole text

is obtained, and the high-dimensional text vector is

reduced. The text vector is mapped to a vector of

length m, where m is the number of label categories.

The corresponding probability of each category is

obtained through a softmax layer, and the calculation

formula is shown in (8). Where W

∈R

m*d

is the

parameter matrix and b

∈R

is the bias parameter. The

event element category corresponding to each word is

shown in (9).

𝑦



= 𝑠𝑜𝑓𝑡𝑚𝑎𝑥

(

𝑊



𝐻





+ 𝑏



)

. (8)

𝑦 = 𝑎𝑟𝑔𝑚𝑎𝑥

(

𝑦



)

. (9)

3 RESULTS AND DISCUSSION

3.1 Experimental Setup

The experimental data set is CEC Corpus. The

number of event elements in the data set is shown in

Table 2. There are 1414 time elements, 1679 place

elements and 5424 object elements.

Table 2: Event Element Labeling In CEC Corpus.

Element

Category

P(%) R(%) F(%)

time element 82.6 79.0 79.1

place element 78.7 71.6 75.0

object element 76.8 70.6 73.6

average 79.4 73.7 76.0

Through the statistical results in Table 4, we can

see that the Chinese event element model proposed in

this paper has achieved good results in CEC Corpus.

The experimental results show that the average F

value of the three Chinese elements is 76%, and the

average accuracy and recall rate are 79.4% and 73.7%

respectively. Among them, all the evaluation

standards of time element have achieved the best

results, and the detection results of object element are

slightly worse than those of event element and place

element, which is related to the fact that the text

feature information of time element and place element

is easier to extract than that of object element.

The experimental results of this method are

compared with those of other methods. The detailed

experimental results are shown in Table 5. It can be

seen from the experimental results that the method

proposed in this paper is second only to the automatic

annotation method in F value, because the method

proposed in this paper is developed from the

perspective of sequential pattern mining, and uses the

artificially constructed rules to extract the event

elements directly, and does not use the information of

event trigger words, so it is a simple sequential

annotation extraction method. Its recall rate has

reached nearly 90%, so the F value is the highest. This

method is better than the traditional dependency

parsing and CRF method. Dependency parsing makes

use of the grammatical association among the

components in a sentence, but it does not make full use

of the semantic association among the words in the

sentence. The method based on CRF obtains the

feature function through training, and uses the feature

function to predict the sequence annotation. This

process does not go deep into the text features in

sentences.

Table 5: Comparison of experimental results.

Methods Adopted P(%) R(%) F(%)

dependency parsing [25] 75.3 71.2 73.2

automatic annotation [41] 74.2 89.6 81.2

CRF [26] 68.1 83.0 74.8

model of this paper 79.4 73.7 76.0

3.3 Analysis of Influencing Factors of

Experimental Results

In order to further study the impact of trigger word

type information and trigger word location

information on Chinese event detection results,

additional experiments are added to explore the

impact of these two factors on the experimental

results. The contrast experiment was divided into four

groups, and the existence of trigger word type

information and trigger word position information

were taken as experimental conditions. In the first

group, two groups of characteristic information were

set as none. In the second group, the trigger word type

information was set to none, and the trigger word

position information was added. The experimental

setup of group 3 was opposite to that of group 2. In

the fourth group, the characteristic information of the

two groups was complete.

Table 6: Experimental test results of influencing factors.

experimental

group No.

Is there

trigger

word type

information

Is there any

trigger word

position

information

F(%)

1 no no 53.2

2 no yes 63.1

3 yes no 65.2

4 yes yes 76.0

The experimental results are shown in Table 6. It

can be seen from the results in the table that when the

trigger word type information and trigger word

position information are used at the same time. The

experimental effect is the best, and the F value is the

highest. Further comparing the F value of the second

group and the third group, we can see that the event

type information is more important than the location

information, and has a greater impact on the

experimental results. It can be seen from the group

An Event Element Extraction Method for Chinese Text

625

with the smallest F value that the experimental effect

is the worst when there is no trigger word type

information and trigger word position information. It

can be seen that the trigger word type information and

trigger word location information are effective word

vector feature information, which further proves the

rationality of selecting these two word vector features

in this model, and optimizes the detection results of

event elements to a certain extent.

4 CONCLUSIONS

This paper mainly introduces the detection method of

Chinese event elements. Using the information of

event trigger words, the type vector and position

vector of trigger words are combined with the pre

training vector and part of speech vector of BERT as

the input of neural network layer. Attention

mechanism is also introduced in the proposed method

to better obtain the association information between

event elements and trigger words. BiLSTM is selected

in the coding layer and decoding layer to calculate the

text feature information. Finally, the proposed event

element detection method and the traditional method

are tested on the CEC Corpus. The analysis of the

experimental results shows that the proposed Chinese

event element detection model has achieved good

results, and can be competent for the task of Chinese

event element detection to a certain extent. At the end

of this chapter, we also discuss the influence of trigger

word type information and trigger word location

information on event element detection results.

In addition, the event extraction models proposed

in this paper are all applied to the event extraction task

at sentence level, and there is no research on the event

extraction from the text level. How to effectively

obtain the information features of the text level

documents and extract the event information from

them is the content of the follow-up work.

ACKNOWLEDGMENT

Supported by 2021 Jiangsu University Brand

Specialty Construction Project Phase II computer

science and Technology Foundation, and Science and

Technology on Information System Engineering

Laboratory (No: 05202004).

REFERENCES

Doddington G R, Mitchell A, Przybocki M A, et al. The

automatic content extraction (ace) program-tasks, data,

and evaluation[C]. The International Conference on

Language Resources and Evaluation, 2004, 2(1): 837-

840.

Xiyue Guo, Tingting He. Review of information

extraction[J]. Computer Science, 2015, 42 (02): 14-17.

Xiaodan Su. Research on Key Technologies of open area

event extraction[D]. Zhengzhou: Information

Engineering University of PLA, 2017.

Yajun Zhang. Research on some key technologies in event

ontology construction[D]. Shanghai: Shanghai

University, 2017.

ICPDI 2022 - International Conference on Public Management, Digital Economy and Internet Technology

626