AN EXTRACTION METHOD OF TIME-SERIES NUMERICAL DATA

FROM ENTERPRISE PRESS RELEASES

Masanori Akiyoshi, Mayu Gen, Masaki Samejima, Norihisa Komoda

Osaka University

Yamadaoka 2-1, Suita, Osaka 565-0871, Japan

Keywords:

Time-series numerical data, enterprise press releases, automatic extraction.

Abstract:

This paper addresses an extraction method of time-series numerical data from enterprise press releases for

business strategy design. Business strategy consists of logical actions for continuously producing enterprise

outcome. The business strategy design process that is partially based on competitive environment analysis may

extremely resort to professional skills so far. To enhance and accelerate the competitive environment analysis,

we focus on press releases of competitors in order to extract numerical data related to products or services.

Sentences in press releases are well organized and grammatically correct. Therefore such extraction is simply

done by identifying the keywords of products or services and the unit indicator co-occurrence. In addition to

such simple approach, we clarify the speciﬁc rules to applying our method to practical press releases.

1 INTRODUCTION

Recently business is considered to be under unpre-

dictable environment, which means rapid changes

of consumers’ demand cause sudden drop of proﬁt

and weaker positioning in the target market. When

present corporate management is in excellent state,

future corporate management is no longer on the elon-

gation of it. To judge investment adequately, it is in-

dispensable to analyze various factors, for instance,

not only ﬁnancial aspects but also aspects of organi-

zational growth, customer satisfaction, and so forth.

On the other hand, along with the development of en-

terprise information systems, it becomes possible to

judge business operations rapidly and adequately, for

instance, sales planning by using OLAP (on line an-

alytical processing) of data warehouse or data min-

ing technique. Even if such business operations are

succeeded, continuous growth of business is not at-

tained without business strategy. Business strategy

has become more critical to corporate management

(Loebbecke, 2003).

Business strategy consists of logical actions for

continuously producing enterprise outcome. Such de-

sign may extremely resort to professional skills so

far, because the design process involves repetitive

work of various analysis, goal setting, and investiga-

tion through simulation, which are inherently human-

centric work. It takes much time to design business

strategy only by hand of human professionals. There-

fore there is a strong need to make it possible to de-

sign business strategy by newly emerging information

technology, which is expected to support human pro-

fessionals.

While there are several possible approaches to sup-

port human professionals, we focus on document

analysis-based support for repetitive analysis. Busi-

ness documents such as enterprise annual reports and

press releases indicate applied results of business

strategy. Market-related documents such as consumer

opinion collected through antenna shops or a callcen-

ter include potential needs for products or services.

Therefore we think analysis on business documents

is inevitable to design business strategy (Akiyoshi,

2005).

In this paper, as analysis on business documents,

we focus on press releases of competitors in order

to extract numerical data related to products and ser-

vices, and so forth. We ﬁrst describe some features

of press releases, and then propose our extraction

method in detail. Finally experimental results are also

discussed.

221

Akiyoshi M., Gen M., Samejima M. and Komoda N. (2006).

AN EXTRACTION METHOD OF TIME-SERIES NUMERICAL DATA FROM ENTERPRISE PRESS RELEASES.

In Proceedings of the First International Conference on Software and Data Technologies, pages 221-225

DOI: 10.5220/0001318202210225

 SciTePress

2 EXTRACTION OF NUMERICAL

DATA FROM PRESS RELEASES

2.1 Approach Based on the Feature

of Press Releases

As mentioned in the previous section, it is indispens-

able to investigate competitors’ status such as the

amount of shipment related to some speciﬁc products

and settlement of accounts in the last quarter, and so

forth. Such numerical data are appeared in press re-

leases, currently via Web pages. Therefore we can

use such Web pages as analysis data source. These

documents have the following features.

• Sentences are well organized and grammatically

correct.

• Similar expression and format is used in the same

contents type of releases.

• Published date is explicitly included.

These features make it possible to ﬁnd the target

portions systematically to some extent, for instance,

shipment data co-occurred with the same keywords

is identiﬁed in the different press releases. Figure 1

shows an overview of the tool based on our proposed

method.

Figure 1: Overview of the extraction tool.

However the contents structure demands different

methods to extract such numerical data from the iden-

tiﬁed target portions in documents, for instance, the

portions are expressed in plain text or in listed items,

and so forth. In addition, even if a portion includes the

target keyword, the extracted numerical data might be

irrelevant for the sake of the description for the other

related keyword. Therefore extraction of numerical

data has to be considered based on such variations

of contents structure and description in documents.

Also extracted pairs of time-stamp and numerical data

are not always complete from graph charts points of

views, because expression on time-stamp varies from

documents, and numerical data at some time-point is

missing. Against these problems, our approach is as

follows.

• Identiﬁcation of the target portions is executed by

both target keywords and prohibited keywords, for

instance, ”card name” as the target keyword and

”card name” indicating electronic money usage as

the prohibited keyword.

• Extraction from the target portions is categorized

into two cases such as plain text portions and listed

item portions.

• Pre-process for the parenthesised portions is used.

• Post-process for the uniﬁed expression of time-

stamp is sometimes used.

• Post-process for ﬁlling up the interval numerical

data is sometimes used.

2.2 Basic Extraction Method

Extraction is processed sentence by sentence. The

keywords and the unit indicator for numerical data

given by a user are input to the tool. Basic extraction

is executed when the keywords except prohibited ones

and the unit indicator co-occurred in a sentence. We

use tagging tool called NExT that automatically iden-

tiﬁes speciﬁc nouns such as person name, organiza-

tion name, location name, and numerical data (Fuku-

moto, 2002). In order to tag arbitrary numerical num-

ber expressions, it is necessary to register unit indica-

tors in its dictionary. Therefore we add online regis-

tration functionality as pre-process using NExT. Fig-

ure 2 shows an example of the basic extraction pro-

cessing ﬂow.

Figure 2: Basic extraction processing ﬂow.

The basic extraction is effective to the sentences

that include only one set of keywords and numeri-

cal data. However, most of sentences include plural

ICSOFT 2006 - INTERNATIONAL CONFERENCE ON SOFTWARE AND DATA TECHNOLOGIES

222

numerical data or optional expressions. Here we in-

troduce processing rules in addition to the basic ex-

traction.

2.3 Extraction by Words Distance

Figure 3 shows an example of plural set of keywords

and numerical data. In this sentence, basic extraction

processing makes two pairs as indicated in Figure .

To distinguish the correct pair from the incorrect, we

introduce words distance rule. As mentioned before,

one of the characteristic features appeared in press re-

leases is ”Similar expression and format is used in the

same contents type of releases”. Therefore the same

criteria of words distance such as ”nearest” is appli-

cable to target sentences. In this example, the ﬁrst

pair indicating time-stamp as ”1l (the ﬁrst quarter)”

and numerical data as ”1,983,739 (1,983,739 thou-

sand yen)” is extracted as correct one.

Figure 3: Example of extraction by words distance.

2.4 Extraction by Pre-process on

Parentheses

Most of additional explanation expressions are ap-

peared in parenthesis-style format such as the use of

round brackets. In such cases, parenthesis portions

play several roles in the sentence. Based on prelim-

inary analysis, we categorize these cases into three

types as follows.

Case A: The parenthesis includes the target key-

words, time-stamp, and numerical data.

Case B: The parenthesis includes time-stamp, and

numerical data except the target keywords.

Case C: Neither case A nor case B occurs.

In the case A, the extraction using tagging is ex-

ecuted on the parenthesis, and extraction is done in

advance. Then, removing the parenthesis and its con-

tents from the sentence is processed for further extrac-

tion.

In the case B, the parenthesis and its contents are

just removed from the sentence for further extraction.

In the case C, only the pair of parenthesis is re-

moved from the sentence.

The above procedures are repeated until all paren-

theses are removed. Figure 4 illustrates an example of

this procedure.

Figure 4: Example of extraction from parenthesis portions.

2.5 Uniﬁed Expression of

Time-stamp as Post-process

As expression on time-stamp, concrete description is

not always used in target documents, for instance,

”the ﬁrst quarter of this year” or ”previous month”

are sometimes used. If such relative description is

intermingled with concrete one, uniﬁed expression is

necessary to generate a set of pairs of time-stamp and

numerical data for generating graph charts.

To unify this expression, one concrete description

is hooked ﬁrst. Then, based on relative description

patterns, the relative description is substituted by us-

ing the hooked concrete time-stamp.

2.6 Filling Up the Interval Data as

Post-process

Suppose that extracted data have the year-based nu-

merical data and the ﬁrst half year-based one as ship-

ping products. Then the second half year-based nu-

merical data is easily calculated from these two data.

In case of generating half year-based graph charts,

this ﬁlling up procedure is executed as post-process

of extraction.

AN EXTRACTION METHOD OF TIME-SERIES NUMERICAL DATA FROM ENTERPRISE PRESS RELEASES

223

3 EXPERIMENTAL RESULTS

Here we discuss experimental results when applying

our method to practical press releases. Three sets of

enterprise press releases are used. Table 1 shows such

experimental conditions; each English meaning is in

round brackets.

Table 1: Experimental conditions.

Number

of docu-

ments

Keywords Unit

indi-

cator

Company-

100 Edy,J[h(card),

s(number of

issues)

Company-

48 Suica,J[h(card),

s(number of

sheets of issues)

Company-

43 A(consolidates),

(sales re-

sults),(sales)

(yen))

Figure 5, Figure 6, and Figure 7 show the result

graphs indicating the retained problem. As shown in

some graphs, some numerical data are not extracted,

and wrong extraction also occurs.

Figure 5: Extracted results on company-A.

As for the missing extraction, it is caused by

slightly different keywords such as ”s (number of is-

sues)” instead of ”s (number of sheets of issues)”. In

our method, a keyword is matched exactly, which re-

lies on the characteristics of press releases ”Sentences

are well organized. Similar expression and format

is used”. Therefore the extension to handle similar

words of input keywords is necessary in more accu-

racy of extraction.

As for the wrong extraction, it is caused by ir-

relevant numerical data co-occurred with an input

keyword in the sentence, for instance, a sentence of

”We commemorate the number of issues over 1000

Figure 6: Extracted results on company-B.

Figure 7: Extracted results on company-C.

as shopping card xxx, the next January 2006 c” in-

cludes ”January 2006” as time-stamp and ”1000” as

numerical data. This is quite difﬁcult without analy-

sis of sentence meaning, so further study is necessary

to handle it.

4 RELATED WORKS

Keywords extraction is frequently discussed in sev-

eral types of documents such as newspapers, aca-

demic papers, Web pages, and so forth. Work in

(Kobayashi, 2005) discusses automatic generation of

keywords map related to technology. TF-IDF (term

frequency and inverse documents frequency) auto-

matically ﬁnds speciﬁc keywords that characterize

documents. ThemeRiver visualize topic stream of on-

line news by using a river ﬂow metaphor (Harve,

2002). However, time-series numerical data extrac-

tion is not discussed enough so far.

Other works intended to extraction of key phrase

or time-series data do not cover the set of ”keyword”,

”time-stamp”, and ”numerical data” (Zhang, 2005;

Sripada, 2003). Of course our used tool NExT han-

dles the extraction of numerical data, but it is general-

use tool. Therefore we investigate what is signiﬁcant

ICSOFT 2006 - INTERNATIONAL CONFERENCE ON SOFTWARE AND DATA TECHNOLOGIES

224

method to extract time-series numerical data for busi-

ness analysis process points of views. In addition,

we discuss post-processes for the sake of generating

graph charts.

5 CONCLUSION

In this paper, we proposed an extraction method of

time-series numerical data from enterprise press re-

lease to enhance and accelerate the competitive envi-

ronment analysis. Our method simply uses the tag-

ging techniques to extract appropriate portions of tar-

get documents. Through experimental results, time-

series data are extracted appropriately, however, the

extension on handling similar keywords is necessary

for more accuracy.

REFERENCES

Loebbecke, C. et al. (2003) The impact of eBusiness and the

information society on ’STRATEGY’ and ’STRATE-

GIC Planning’: An assessment of new concepts and

challenges. J. of Information Technology and Man-

agement, vol.4, pp.165-182

Kobayashi, S. et al.(2005) Time series analysis of technol-

ogy trends based on the Internet resources IEEJ trans.

On EIS, vol.125, no.5, pp.720-729

Harve, S. et al.(2002) ThemRiver: Visualizing thematic

changes in large document collections IEEE trans.

On Visualization and Computer Graphics, vol.8, no.1,

pp.9-20

Akiyoshi, M. et al.(2005) An analysis framework of enter-

prise documents for business strategy design In Proc.

of 5th Int. Conf. on Computational Intelligence for

Modelling Control and Automation (CIMCA’2005),

vol.2, pp.330-335

Fukumoto, J. et al.(2002) A named entity extraction tool

(NExT) for text processing (in Japanese) In the Proc.

of 8th Annual Meeting of The Association for Natural

Language Processing, pp.176-179

Zhang, Y. et al.(2005) Narrative text classiﬁcation for au-

tomatic key phrase extraction in web document cor-

pora In Proc. of the 7th annual ACM international

workshop on Web information and data management,

pp.51-58

Sripada, S. G. et al.(2003) Generating English summaries

of time series data using the Gricean maxims In Proc.

of the 9th ACM SIGKDD international conference on

knowledge discovery and data mining, pp.187-196

AN EXTRACTION METHOD OF TIME-SERIES NUMERICAL DATA FROM ENTERPRISE PRESS RELEASES

225