Time Phrase Parsing for Chinese Text with HowNet Temporal

Information Structure

Hong-mei Ma, Xiao-yun Wang and Li Qin

Library, China Defense Science and Technology Information Center, Beijing, China

Keywords: Information Retrieval, Temporal Information, Time Phrase Parse, Chinese Text.

Abstract: Time phrase parsing is useful to analyze Chinese text, because it is usually omitted expression in the

Chinese text for the coherence and the cohesion. It is important to gain the temporal information, then it can

be used to generate the tense of the verb. In this paper, with temporal information structures of HowNet,

time phrases are divided into three categories, and a structure is designed to represent the temporal

information of time phrases. Then, it puts forward a method to parse time phrases in Chinese text.

1 INTRODUCTION

As we all know, there isn’t morphologic changing in

Chinese. Temporal informal implies in the Chinese

text. Since temporal system is objective to every

kind of natural language, we could gain the temporal

information from the Chinese text. It has been

proved that the temporal system has three

components, phase, tense and aspect (Zhou 2000).

Phase is a property of the event described by the

semantic of the verb, tense is a temporal relation

between the event time (ET) and the speech time

(ST) or another reference time (RT), and aspect is

the state of the event, which depends on the

point-of-view of the observation (M & G 1996).

Temporal relations are precedence, simultaneous and

include-in. Note that the tense discussed in the paper

is different from the syntactic category tense in

English, values of it are happen-ed, happen-ing and

happen-will. Happen-ed represents ET is precedence

to ST (or RT), happen-ing means ET is simultaneous

to ST (or RT), or ET is included in RT, happen-will

indicates ST (or RT) is precedence to ET. In the

Chinese language time phrase is a syntactic approach

to represent the tense of the verb (Gong 1995), it

often has temporal references among the sentences

in the Chinese text for the coherence and the

cohesion, so it is necessary to parse the time phrases

in the Chinese text in order to determine the tenses

of the verbs in the text.

HowNet is a bilingual general knowledge-base

describing relations between concepts and relations

between the attributes of concepts. It has more than

2,000 sememes to define the concepts and proposes

information structures to reveal the meaning of

phrases at the semantic level (Dong 2007). Chinese

information structures database of HowNet (short for

CISD) is derived from large scale natural language

corpus, which gives 268 information structures and

more than 10,000 examples in the latest version

(Dong 2010). In CISD there are 23 temporal

information structures and 1,029 examples of time

phrases, then we analyse these temporal information

structures and experiment with the examples. In

order to computing semantic of time phrases, we

category time phrases and design a semantic

structure, named TPSS, to represent the temporal

information of time phrases.

In the paper, time phrases are categorized into

three in section 2, then, a method is put forward to

parsing time phrases in section 3. Finally, we

implement a system to parsing time phrases in

Chinese texts.

2 TIME PHRASE

CATEGORIZATION

Constituents of time phrases include time noun

words, for example, 昨天(zuo2 tian1, yesterday),

time expression that composed of number and

time-quantifier, such as 2013 年 5 月(er4 ling2 yi1

san1 nian2 wu3 yue4, May 2013), in which 年

294

Ma H., Wang X. and Qin L..

Time Phrase Parsing for Chinese Text with HowNet Temporal Information Structure.

DOI: 10.5220/0005094902940299

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2014), pages 294-299

ISBN: 978-989-758-048-2

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

(nian2, year) and 月 (yue4, month) are two

time-quantifiers, time order words, for example, 前

(qian2, preceding), and part-of time words, for

instance, 初(chu1, beginning) in Chinese. Semantic

of time phrases is illustrated by temporal information

structures of HowNet (short for TIS) intensively, for

example:

(1) 昨天-上午(zuo2 tian1 shang4 wu3, yesterday

morning)

TIS=(时间|time,特|special,日|day) [whole] 

( 时间|time, 特 |special, 晨 / 午 / 晚

|morning/afternoon/night)

(2) 会-前(hui4 qian2, before the meeting)

TIS=(事情|things)  [time] (时间|time)

In (1) it describes the morning in yesterday,

concept yesterday is the whole of concept morning.

Example (2) is about the time before the time when

the meeting is began. According to the meaning of

time phrases, we category temporal information

structures of HowNet and time phrases into three

kinds, date, presudodate and time interval.

2.1 Date

Date is the numbers or words used to talk about a

particular day, month, and year. It has the definite

value on the axis of time. The formation of the time

phrase expressing the date in Chinese is shown as

follows:

Date::= [<Time-unit>]* (1)

Time-unit::= <time-num><time-quantifier> (2)

The time-quantifiers of the dates include 世纪

(shi4 ji4, century), 年(nian2, year), 季(ji4, season),

月(yue4, month), 旬(xun2, ten days), 周(zhou1,

week), 日(ri4, day), 日段(ri4 duan4, part of a day),

时(shi2, hour), 分钟(fen1 zhong1, minute) and 秒

钟 (miao3 zhong1, second) (Dong 1999). The

time-unit that composed of the former

time-quantifier is larger than the time-unit composed

of the latter time-quantifier. Moreover, 季 has 4

values: 春(chun1, spring), 夏(xia4, summer), 秋

(qiu1, autumn) and 冬(dong1, winter), which have

the precedence relation. And 日段 also has 4 values:

晨(chen2, morning), 午间(wu3 jian1, noon), 下午

(xia4 wu3, afternoon) and 晚(wan3, night), which

also have precedence relation. Finally, every

time-unit of the dates could be divided into three

parts again, which are the beginning part, the middle

part and the ending part, and there is precedence

relation between these parts.

The dates have two properties in the Chinese text.

One is that the dates often only contain the larger

time-unit when it isn’t necessary to point out the

detail time. The other is that the dates often have the

temporal references in the Chinese text. The larger

time-unit of the date would be omitted when it is

same as the larger time-unit of the date in the above

sentences.

2.2 Presudodate

Presudodate is composed of a based-date and an

offset, in which the based-date is a date, and the

offset consists of direction and size. Because the

offset and the based-date of the presudodate are

often indefinite, the presudodate hasn’t definite

value on the axis of time. However, the presudodate

could be transformed to the date when the base-date

and the offset are given. Presudodate is defined as

follows:

Presudodate ::= <Based-date><Offset> (3)

Offset ::= <Offset-dir><Offset-size> (4)

Date=Presudodate.Based-date + Presudodate.Offset

(5)

Expression (5) shows the process of transforming

the presudodate to the date. When the direction of

the offset is same as the direction of the axis of time

(the flow of time), the size of the offset is positive.

Well, the size of the offset will be negative when the

direction of the offset is opposite to the direction of

the axis of time. And when the size of the offset is

zero, it implies that the presudodate denotes the

based-date, or the larger time-unit of the presudodate

and that of the based-date are same, for example,

今

天

(jin1 tian1, today),

今天下午

(jin1 tian1 xia4 wu3,

this afternoon).

Time phrases expressing the presudodates have 4

forms generally in the Chinese text:

(1) Have the based-date and the offset direction. For

example, 2008

年

月

日前

(er4 ling2 ling2 ba1

nian2 liu4 yue4 yi1 ri4 qian2, before June 1st

2008). This kind of presudodate could be

transformed to the date with formula (5) after

receiving the size of the offset.

(2) Have the offset including the direction and the

size, and take the latest date in the above

sentences as its based-date. For example,

“1999

年，…。两年后，…。”

(yi1 jiu3 jiu3 jiu3

nian2, …liang2 nian2 hou4,…，… in 1999. …

two years later …).

(3) Have the offset, and take the date that the event

happened as its based-date. For example,

“第二

次世界大战以后”

(di4 er4 ci1 shi4 jie4 da4

TimePhraseParsingforChineseTextwithHowNetTemporalInformationStructure

295

zhan4 yi2 hou4，after The Second World War ).

(4) Have the offset, and take the date that the text

was written as its based-date. For example,

“明

天”

(ming2 tian1, tomorrow), its offset direction

is positive and its offset size is one day.

2.3 Time Interval

What the third kind of time phrases represents is a

time interval, which has two different forms in the

Chinese text. One is composed of the beginning date

and the ending date, such as 80

年代末

年代初

(ba1 shi2 nian2 dai4 mo4 jiu3 shi4 nian2 dai4 chu1,

from the ending of 1980s to the beginning of 1990s),

the other just represents the size of the time interval,

for instance,

十天

(shi2 tian1, ten days). By the way,

the size of the offset of the presudodate mentioned

above is a time interval.

3 TIME PHRASE PARSING

3.1 TPSS

TPSS is the abbreviation for the Time Phrase

Semantic Representation Structure. It contains the

temporal information of a time phrase, its definition

is:

TPSS ::= <Semcate, Semcontent> (6)

Semcate  {date, presudodate-basedate,

presudodate-timeinterval}

Semcontent ::= <Date, Offset-dir, Offset-size> (7)

Offset-dir::= nega | zero | posi (8)

Offset-size ::= [<offset-size-unit>]*|<how-adv> (9)

offset-size-unit ::= <time-num>

[<special-quantifier>]<time-quantifier> (10)

Semcate is the category of the time phrase that

we have discussed in section 2. Semcontent is the

temporal information of the time phrase defined as a

triple including date, offset-dir and offset-size.

How-adv is the adverb representing the degree of the

quantity, for instance, 左右(zuo3 you4, or so), 整

(zheng3, exactly). And the special-quantifier is some

quantifier that could be used between the number

and the time-quantifier.

TPSS is an enclosure of the temporal information

of the date, the presudodate and the time interval.

When the Semcate of TPSS is date, the offset-dir

and the offset-size is zero. When TPSS represents a

presudodate with the based-date, date of the

Semcontent is the based-date of the presudodate.

And when the Semcate of TPSS is

presudodate-timeinterval, there would be two dates

in TPSS, one is the beginning date, the other is the

ending date, and the size of the time interval would

be represented by the offset-size.

3.2 Mliurn

Linguistic context of Chinese text is important to the

Chinese text understanding. HowNet gives the static

general knowledge about the world, but the dynamic

context knowledge about the text should also be

collected in the process of text analysis. In the paper,

a structure named MLIURN ( Multilevel Information

Unit Relation Network) is designed to represent the

context knowledge of the Chinese text, see Figure 1:

concept

sentence

paragraph

chapter

text

Figure 1: Multilevel Information Unit Relation Network.

As indicate in Figure 1, MLIURN has 5 levels,

every level has information units and the relations.

On one hand, information unit at the low level is part

of information unit at the high level. On the other

hand, information units and relations at different

level have different contents that representing the

knowledge of natural language from different level.

For example, at the concept level, information unit is

one concept which has syntactic properties, semantic

properties and the number of the sentence in which

the concept is contained. The relation between two

information units is the semantic relation between

the two concepts. Concepts expressed in the concept

level include event concepts, entity concepts and

time concepts. TPSS is one information unit at the

concept level. Well, information unit at the text level

describes the information of the text, such as the text

structure, the text category and the theme of the text.

3.3 Time Phrase Parser

Time phrase parser is a Finite-State Machine

including a table of input symbol sets and a function

of state transition. Every state of the parser is an

object with special semantic action. First, it writs the

temporal information of the time phrase into TPSS.

Second, it would to search MLIURN to receive the

KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval

296

temporal information from the above paragraphs

when there are time references in the Chinese text.

The structure of the parser is showed in Figure 2.

Chinese Text Sentences

Time

Phrase

Parser

MLIURN

State Transition

Function

Input Symbol

Sets Table

TPSS

Figure 2: Chinese Text Time Phrase Parser.

We define 18 kinds of input symbol sets as

indicated in Table 1.

Table 1: Input Symbol Sets Table.

A Chinese number set

i.e. “五”(wu3, five)

Month set i.e. “三

月 ”(san1 yue4,

March)

Digit set i.e. “8”

K Presudodate time

word set i.e. “ 昨

天 ”(zuo2 tian1,

yesterday)

C Time-quantifier set

i.e. “ 月 ”(yue4,

month)

L Presudodate prefix

set i.e. “ 过

去 ”(guo3 qu4,

bypast)

D Time pronoun set i.e.

“那”(na4, that)

M Presudodate suffix

set i.e. “以前”(yi3

qian4, before)

PartofDay set i.e “上

午”(shang2 wu3, in

the morning)

N Special quantifier

set i.e. “个”(ge4)

F Year Prefix set i.e.

“公元”(gong1 yan2)

O Time order word

set i.e. “第”(di4)

G Quantity adv set i.e.

“ 左右”(zuo3 you4,

or so)

P Other time word set

i.e. “国庆节”(guo2

qing4 jie2, National

Day)

H Part-of time word set

i.e. “ 初 ”(chu1,

beginning)

Q Education place set

i.e. “ 大学”(da4

xue2, university)

Weekday set i.e. “星

期一”(xing1 qi1 yi1,

Monday)

Verb set i.e. “ 毕

业 ”(bi4 ye4,

graduate)

Time phrase parsing algorithm is as follows:

Time Phrase Parsing Algorithm（TPPA）

Let S

be the current state of the parser, X be

one set of Table 1, and F be the state transit function.

1) Input word w.

2) Evaluate the word w whether or not

belong to one set of Table 1.

3) If not, goto 9).

4) If w  X, calculate the state S’ to

which S

transit by F(S

, w).

5) If S’=NULL, return to the last

termination state, goto 9).

6) If S’NULL, set S’ to S

7) Perform the semantic action of S

8) If S

is a termination state，write the

result of 7) to TPSS, goto 1).

9) If TPSS=NULL, return;

10) If TPSSNULL, test TPSS whether or not

having temporal reference to the above

paragraph.

11) If there is temporal reference, search

MLIURN, extend TPSS with the temporal

information of the time phrase in the

above paragraph.

12) Write TPSS into MLIURN, return.

3.4 Tense Calculus

Now the tense of the verb that restricted by the time

phrase could be calculated. First, some simple

operators are defined, and then an algorithm

( Chinese Text Verb Tense Parsing Algorithm) to

calculate the tenses of the verbs in Chinese text.

DEFINITION 3.1 Date plus operator +

Let X be a set of dates, and Y be a set of offsets,

date plus operator + is a transformation from (X, Y)

to X .

Suppose



d  X，



o  Y ，

d = {time-unit

, time-unit

, … , time-unit

o = {offset-size-unit

, offset-size-unit

, … ,

offset-size-unit

}, m, n Z.

d + o =

d.time-unit

.time-num +

o.offset-size-unit

.time-num iff

d.time-unit

.time-quantifier =

o.offset-size-unit

.time-quantifier,

i ,j {1,2,…,n}

EFINITION 3.2 Date minus operator -

Let X be a set of dates, and Y be a set of offsets,

date minus operator - is a transformation from (X, Y)

to X .

Suppose



d  X，



o  Y ，

d = {time-unit

, time-unit

, … , time-unit

o = {offset-size-unit

, offset-size-unit

, … ,

offset-size-unit

}, m, n Z.

d - o =

d.time-unit

.time-num –

o.offset-size-unit

.time-num iff

TimePhraseParsingforChineseTextwithHowNetTemporalInformationStructure

297

d.time-unit

.time-quantifier =

o.offset-size-unit

.time-quantifier ,

i ,j {1,2,…,n}

DEFINITION 3.3 Date compare operator R

Let X be a set of dates, and Y be a set of relations

between dates, Y={pre, equ, aft}, date compare

operator R is a transformation from (X, X) to Y .

Suppose

 d

, d

 X,

= {time-unit

, time-unit

, … , time-unit

= {time-unit

, time-unit

, … , time-unit

time-unit

is larger than time-unit

i+1

i=1,2,…n-1.

i) If d

.time-unit

.time-num <

.time-unit

.time-num,

R(d

, d

)= pre .

ii) If d

.time-unit

.time-num >

.time-unit

.time-num,

R(d

, d

)= aft .

iii) If d

.time-unit

.time-num =

.time-unit

.time-num, i = i+1;

continue i), ii), iii).

iv) For i = 1,…n,

if d

.time-unit

.time-num =

.time-unit

.time-num,

R(d

, d

)= equ.

DEFINITION 3.4 Datelization operator D

Let X be a set of dates, and Y be a set of

presudodates, F be a set of the directions of the

offset, F={posi, nega, zero}, datelization operator D

is a transformation from X to Y .

Suppose

 d  X,

i) if d.offsetsir = posi ,

D(d ) = d.basedate + d.offsetsize,

ii) if d.offsetsir = nega ,

D(d ) = d.basedate – d.offsetsize.

In i), ii), d.basedate is the based-date of the

presudodate, d.offsetdir is the direction of the offset

of the presudodate, and d.offsetsize is the size of the

offset of the presudodate.

DEFINITION 3.5 Offset direction operator



Let X be a set of dates, and F be a set of the

directions of the offset, F={posi, nega, zero}.

Suppose

 d

, d

 X,

i) if R(d

, d

)= pre ,



, d

)= nega,

ii) if R(d

, d

)= equ ,

 (d

, d

)= zero,

iii) if R(d

, d

)= aft ,

 (d

, d

)= posi.

 (d

, d

) indicates the direction of the offset

that d

relative to d

EFINITION 3.6 Tense operator T

Let X be a set of dates, and S be a set of the

tenses, S={ happen-ed, happen-ing, happen-will }.

Suppose



, d  X,

i) if



, d)= nega,

T(d

, d)= happen-ed,

ii) if



, d)= zero,

T(d

, d)= happen-ed,

iii) if



, d)= posi,

T(d

, d)= happen-ed,

In i), ii), iii), d is the date when the text is written,

so T (d

, d) indicates the tense of d

which relative to

The algorithm is defined as follows:

Chinese Text Verb Tense Parsing Algorithm

Let d be the date when the text is written, S be

the current sentence, and V be the verb in S.

Step1 If S has time phrase, receive its

semantic structure named TPSS

TPPA.

(1) If TPSS

is a date, denoted d

, compute



, d), and then compute T(d

, d).

(2) If TPSS

is a presudodate, denoted pd

i) If pd

has based-date,

a. If pd

.based-date=d, infer the

tense of pd

from pd

.offset-dir.

b. If pd

.based-dated, compute

D(pd

)= d

, then compute  (d

, d),

finally compute T(d

, d).

ii) If pd

has offset, and take the last

date in the above paragraph as its

based-date, get the date from

MLIURN, then perform as i).

iii) If pd

take the date that the event

happened which in the above

paragraph as its based-date, get the

date from MLIURN, then perform as i).

(3) If TPSS

is a time interval,

i) If TPSS

has the beginning date d

and

the ending date d

, compute  (d

, d)

and



, d). If  (d

, d)= posi,

the tense is happen-will, if

 (d

d)= nega, the tense is happen-ed,

otherwise, the tense is happen-ing.

ii) If TPSS

is a time interval with the

size of the interval, take the tense

of the last sentence in the above

paragraph as its tense.

Step2 If V is restricted by TPSS

, set the

tense of TPSS

to the tense of V.

Step3 If S has no time phrase，take the tense

of the last sentence in the above

paragraph as the tense of V.

KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval

298

4 EXPERIMENTS

We have built a text parsing model and a text context

knowledge representation structure MLIURN in

order to solve some problems of the sentence parsing

model. With the parser, it has parsed the 23 temporal

information structures of HowNet and the examples,

analyzed time phrases in the Chinese text, and

evaluated the tenses of the verbs.

The following text is taken from the newspaper

BEIJING YOUTH DAILY (16 April 2014), and the

analysis result is showed in Table 2.

李嘉诚“七连抛”躲避资产风波

本报讯

近日，李嘉诚旗下的和记黄埔有限公司得到

伦敦市长的许可，获批在伦敦东南部的德特福德兴建房

地产项目。而本月初，李嘉诚次子李泽楷以

9.28

亿美

元出售其掌控的盈大地产旗下北京盈科中心。至此，从

去年

月开始，李家在内地和香港已上演“七连抛”，

“一抛一入”之间引发了市场的相关猜测。据统计，

2010

年至今李嘉诚大举抛售香港和内地资产，其市值已超过

总身价七成。

摘摘自《北京青年报》〔

20140416

〕

Table 2: Parsing Result.

Event Time phrase

Temporal

reference

Tense*

得到(de2 dao4) 近日(jin4 ri4)

happen-ed

获批(huo4 pi1)

Yes happen-ed

兴建(xing1

jian4)

Yes happen-ed

出售(chu1

shou4)

本月初(ben3

yue4 chu1)

happen-ed

上演(shang4

yan3)

从去年 8 月开始

(cong2 qu4 nian2

a1 yue4 kai1

shi3)

happen-ed

引发(yin3 fa1)

Yes happen-ed

抛售(pao1

shou4)

2010 年至今(er4

ling2 yi1 ling2

nian2 zhi4 jin1)

happen-ed

超过(chao1

guo4)

Yes happen-ed

*Tense is one element of temporal system, which isn’t the

syntactic category tense in English. The value indicates the

temporal information of the events in the Chinese text.

5 CONCLUSION

In this paper, temporal information structures of

HowNet and time phrases are divided into 3 kinds,

date, presudodate and time interval. Then, a structure,

called TPSS, is designed to represent the temporal

information of time phrases. Finally, a parser is built

to parse time phrases and to infer the tenses of the

verbs in Chinese texts.

Although the parsing algorithm of time phrases

in Chinese text has shown satisfactory result, some

difficulties are still remained, such as ambiguous

word segmentation. Because of lacking for

commonsense knowledge, tenses of some

presudodates couldn’t be calculated. We will do

more work about the Chinese temporal system in the

future.

ACKNOWLEDGEMENTS

We gain many ideas in the discussion with Mr. Dong

Zhendong. This work would not be possible if

without his advice. We would like to thank him.

REFERENCES

Qianyan Gong, 1995. Phase Tense and Aspect in Chinese,

The Commercial Press.

Gagnon M., and Lapalme G., 1996. From Conceptual Time

to Linguistic Time, Association for Computational

Linguistics.

Chongli Zhou, 2000. Research of Natural Language Logic,

Beijing University Press.

Zhendong Dong, 2007. Theoretical findings of HowNet,

Journal of Chinese Information Processing, vol.21,

no.4, 3-9.

Qiang Dong, 2010. A HowNet-Based Disambiguator for

Chinese Syntactic Structures, Journal of Chinese

Information Processing, vol.24, no.1, 60-64.

HowNet. http://www.keenage.com.

TimePhraseParsingforChineseTextwithHowNetTemporalInformationStructure

299