TERM KNOWLEDGE ACQUISITION USING THE STRUCTURE

OF HEADLINE SENTENCESFROM INFORMATION

EQUIPMENTS OPERATING MANUALS

Makoto Imamura, Yasuhiro Takayama

Information Technology R&D Center, Mitsubishi Electric Corporation

5-1-1 Ofuna, Kamakura, Kanagawa, 247-8501, Japan

Masanori Akiyoshi, Norihisa Komoda

Graduate School of Information Science and Technology, Osaka University

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

Keywords: Linguistic knowledge acquisition, case frame, operating manual, document structure.

Abstract: This paper proposes a method for automatically extracting term knowledge such as case relations and IS-A

relations between words in the headline sentences of operating manuals for information equipments. The

proposed method acquires term knowledge by the following iterative processing: the case relation extraction

using correspondence relations between surface cases and deep cases; the case and IS-A relation extraction

using compound word structures; the IS-A relation extraction using correspondence between the case

structures in the hierarchical headline sentences. The distinctive feature of our method is to extract new case

relations and IS-A relations by comparison and matching the case relations extracting from the super and

sub headline sentences using the headline hierarchy. We have confirmed that the proposed method has

achieved 92.4% recall and 96.8% precision for extracting case relations, and 93.9% recall and 89.9%

precision for extracting IS-A relations from an operating manual of a car navigation system.

1 INTRODUCTION

As the functions of information equipments increase

diversity and complexity in recent years, users can

hardly find the target items from several hundred

pages of operating manuals. Especially for

information equipments such as car navigation

systems, mobile phones, and DVD recorders, the

help of operation is important because the target

users of those are broad range of people (Waida,

2007). To facilitate the help of equipment operations

and manual search, the function of refining user’s

intention is desired, and term knowledge is required

to implement the function (Kurohashi, 2000),

(Kawahara, 2003). The representative term

knowledge are case relations like equipment

operations and their objects, and the super-sub

relations (IS-A relations) between objects of the

equipments. For example, by using term knowledge

of case relations and IS-A relations, help systems

can suggest the related terms with the user-input

keywords to improve search ability. However, the

human workload of term knowledge creation with

expertism for each special domain is at issue.

Therefore, the needs of term knowledge acquisition

from operating manuals, necessarily with the

commercialization of the equipments, are growing.

Conventional technologies of acquisition of case

relations and IS-A relations are based on statistical

machine learning (Oishi, 1997), (Kawahara, 2000)

or sentence pattern rules (Kurohashi, 1992),

(Kobayashi, 2004). The former methods require

huge corpus. The targets of latter methods are

restricted in dictionaries and blogs. Therefore, both

methods cannot be applied to term knowledge

acquisition from operating manuals whose volumes

are relatively small. Moreover, the conventional

methods are based on one sentence processing, so

the methods using the document structure such as

the hierarchy of headlines or the relations between

the cells in the tables, specific to the business

documents, has not been examined.

We focus on the headline hierarchy as the

document structure in the business documents. This

301

Imamura M., Takayama Y., Akiyoshi M. and Komoda N. (2008).

TERM KNOWLEDGE ACQUISITION USING THE STRUCTURE OF HEADLINE SENTENCESFROM INFORMATION EQUIPMENTS OPERATING

MANUALS.

In Proceedings of the Third International Conference on Software and Data Technologies - PL/DPS/KE, pages 301-308

DOI: 10.5220/0001875803010308

 SciTePress

paper proposes a method of automatically acquiring

case relations and IS-A relations, whose distinctive

feature is the matching the case elements by using

the headline hierarchy.

2 TERM KNOWLEDGE OF

ACQUISITION

In the operation help and the manual search for

information equipments, the system can support the

users of the equipments to find the desired

explanation pages quickly by suggesting the related

terms of the user-input keywords using term

knowledge. For example, the user can easily reach

the targeted page by only choosing the desired terms

from the alternatives suggested by the system using

term knowledge shown in Table 1, instead of

wandering around looking for a number of the pages

in the search result list.

Table 1: Examples of term knowledge.

Ex.

{verb/settei-suru (setting),

object/mokuteki-chi (destination place),

instrument/navi-meyu (navigation menu)}

{verb/settei-suru (setting),

object/mokuteki-chi (destination place),

instrument/ touroku-chi (registered place)}

{verb/kensaku-suru (search),

object/ mokuteki-chi (destination place),

instrument/jusho (address)}

Ex.

IS-A( kisetsu-no-supotto (seasonal spot),

osusume-jouhou (recommended

information))

IS-A( hyakusen(the hundred best),

osusume-johou (recommended information))

IS-A(osusume-supotto (recommended spot),

osusume-jouhou (recommended

information))

Ex.1 in Table 1 are called the case frames that

represent the verb centred relations about the role

(deep case) of the noun to the verb. For example,

{verb/settei-suru (setting), object/mokuteki-chi

(destination place), instrument/navi-menyu

(navigation menu)} represents “noun ‘mokuteki-chi

(destination place)’ has a role of ‘object (objective

case)’, and noun ‘navi-menyu (navigation menu)’

has a role of ‘instrument (instrumental case)’ for the

verb ‘settei-suru (setting).’” The case particles with

nouns that depend on the verbs are called ‘surface

cases’. For example, “tourokuchi-wo (registered

place)” is called “wo-case”, “risuto-ni (to the list)” is

called “ni-case”. From Ex.1 in Table 1, we can see

the following: “The verbs with ‘mokuteki-chi

(destination place)’ as the object are ‘settei-suru

(setting)’ and ‘kensaku-suru (search)’”; “The

instruments for ‘mokuteki-chi-wo settei-suru (setting

the destination place)’ are ‘navi-menyu (navigatioin

menu)’ and ‘touroku-chi (registered place).’”

Ex.2 in Table 1 are examples of IS-A relations

extracted from the super-sub relations between terms.

These examples represent: “there are 3 instances for

‘osusume-jouhou (recommended information)’:

‘kisetsu-no supotto (seasonal spot)’ ‘hyaku-sen

(the hundred best)’, and ‘osusume-supotto

(recommended spot).’”

3 RELATED WORK

The methods for automatic extraction of term

knowledge are mainly the named entity extraction

method (Sudo 2001) and the corpus-based

knowledge extraction method (Utsuro 1998). These

methods are classified broadly into the machine

learning based methods and the linguistic structure

patterns based methods.

The machine learning based methods require the

huge workload for preparation, because they require

huge corpus like hundreds of millions of sentences

or the training data created by human. Then, there

are the problems: (1) It is difficult to collect huge

corpus for certain business domains. (2) The cost for

creating answer data is high. Therefore, it is hard to

apply the methods to acquire technical term

knowledge in business domains.

The linguistic structure patterns based methods

enable term knowledge acquisition to decrease the

required size of documents by using the human-

created pattern rules specified in each type of

documents. The examples of these methods are

“term knowledge extraction of hyper text

construction from dictionary (Kurohashi 1992)”, and

“term knowledge acquisition for sentiment

extraction (Kobayashi 2004).” The former specialize

the target documents for extraction in the dictionary

and the latter specialize those in Web bulletin boards

and blogs. Because the pattern rules need to be

created for each type of documents, the specific

pattern rules are difficult to apply to the other type

of documents. We call this issue “issue 1: using

characteristics of the headline sentences in the

information equipment manuals.”

Moreover, the conventional methods are based

on processing one sentence, so the case relations and

IS-A relations acquisition using the headline

hierarchy has not been examined. We call this issue

“issue 2: case relations and IS-A relations in the

ICSOFT 2008 - International Conference on Software and Data Technologies

302

headline hierarchy.” The conventional methods use

the surface dependency relations with case particles,

so extracting the case relations between words that

consist of the compound words has not been

examined. We call this issue “issue 3: case relations

in the compound words.”

4 TERM KNOWLEDGE

ACQUISITION USING THE

HEADLINE STRUCTURE

4.1 Overview of Proposed Method

To cope with “issue 1: using characteristics of the

headline sentences in the information equipment

manuals”, we propose a method of extracting case

relation by the correspondence rules between surface

and deep case based on the speciality of linguistic

expressions of operations, status and conditions in

the operating manuals. We also propose a method of

extracting IS-A relation by using the commonality of

word ending in compound words.

To cope with “issue 2: case relations and IS-A

relations in the headline hierarchy”, we propose a

method of extracting case relation by merging cases

in the super and sub headlines when the both verbs

are identical, but cases are different. We also

proposes a method of extracting IS-A relations, by

comparing with cases in the super and sub headlines

when the both verbs are the identical, and both cases

are also identical.

To cope with “issue 3: case relations in the

compound words”, we propose a case analysis

method for compound words by using the linguistic

expressions specific in the operating manuals.

Results of

syntactic analysis

for each sentence

Hierarchy of

headline sentences

Case frame

dictionary

IS-A relation

dictionary

surface case analysis

common ending word processing

Different type case merging

compound word case analysis

Same type case refinement

Case relation extraction processing group

IS-A relation extraction processing group

Figure 1: The architecture of the automatic term

knowledge acquisition method.

The architecture of our proposed term knowledge

acquisition consists of the processings corresponding

to the above methods, the case frame dictionary that

is a set of case relations, and the IS-A relation

dictionary that is a set of IS-A relations as shown in

Figure 1. Each processing extracts new case

relations or IS-A relations by using the results of

syntactic analysis for each sentence, the hierarchy of

headline sentences, the case frame dictionary, and

the IS-A relation dictionary, then registers the new

relation in the case frame dictionary or the IS-A

relation dictionary. The whole process of the

proposed method is repeated iteratively until no new

term knowledge can be extracted.

4.2 Extraction of Case Relations

4.2.1 Case Relations

There is no consensus on what deep cases exist

(Tsujii 1985), therefore, we use the deep cases

shown in Table 2, which are customized from

Fillmore’s deep cases (Fillmore 1975) for the

operations of information equipments.

Table 2: Deep cases for information equipment operations.

Deep case Explanation

Agentive case (A) Information equipment of

operation target, or the parts of it

Experiencer case (E) User of the equipment

Objective case (O) Operation target of the users

Instrumental case (I) Instrument or mean of equipment

operation

Source case (S) Status or place before equipment

operation

Goal case (G) Status or place after equipment

operation

Locative case (L) Place or location about equipment

operation

Time case (T) Condition of equipment operation

Table 3: Examples of case relations in the headlines.

Headline sentence Case relation

touroku-chi-no sentaku

(choice of the registered

place)

{verb/sentaku-suru (choose),

object/touroku-chi

(registered place)}

koe-niyoru sousa

(operation by voice)

{verb/sousa-suru (operate),

instrument/koe (voice)}

Saisei-gamen-kara navi-

gamen-ni modosu

(bring replay screen back

to the navigation screen)

{verb/modosu (bring back),

source/saisei-gamen

(replay screen),

goal/navi-gamen

(navigation screen)}

Table 3 shows the examples of case relations

obtained by analysing the headline sentences in the

operating manuals of a car navigation system. Here

we denote the cases by attribute-value pairs whose

attribute names are the deep cases and values are the

words or phrase in the sentences. For example, {verb

/sentaku-suru (choose), object/touroku-chi

TERM KNOWLEDGE ACQUISITION USING THE STRUCTURE OF HEADLINE SENTENCES FROM

INFORMATION EQUIPMENTS OPERATING MANUALS

303

(registered place)} represents “the objective case of

the verb sentaku-suru (choose) is touroku-chi

(registered place).”

4.2.2 Surface Case Analysis

The corresponding relations between surface and

deep cases become simple when we restrict the

sentences to the headlines of operating manuals for

equipments. The processing for the correspondence

between surface and deep case extracts the case

relations by determining deep cases from the

analyzed results of each sentence using the rules

shown in the left and the upper-right column in

Table 4. The underlined expressions in Table 4 are

examples of surface cases.

In Table 4, the operation words, status words,

and conditional expressions, specific to the operating

manuals of information equipments, are as follows.

(i) Operation Words. The operation words are the

verbs and sahen-nouns in the headline sentences.

For example, “settei (setting)” in “mokuteki-chi-no

settei (setting of the destination place)” and “meiru

(mail)” in “meiru kinou (mail function)” are the

verbs, and then they are the operation words.

(ii) Status Words. The status words are nouns that

are related by the verbs with the expressions of

beginning, duration, and completion. For example,

“sukurouru-suru (scroll)” is an operating word,

auxiliary verb “ta (have -ed)” represents completion

of (9) in Table 4: “sukurouru-shita chiten (the place

that have been scrolled to)”, so the noun “chiten (the

place)” is a status word. The examples of beginning,

duration, and completion are “hajimeru (start)”,

“teshimau (have -ed)”, and “teiru (have been)”. We

can make the dictionary beforehand for the status

words, because these expressions are formulaic and

domain independent.

(iii) Conditional Expressions. The conditional

expressions are terms that represents conditions such

as “baai (case)”, “ji (time)”, and “sai (as)”. The

conditional expressions have functions as the surface

cases. For example, the “ji (time)” in “setsuzoku-ji-

no kakunin (confirmation at the time of connection)”

of (10) in Table 4 is a conditional expression, and

“ji-no (at the time of connection)” has a function of

time case. We can make the dictionary beforehand

for the condition words, because these expressions

are formulaic and domain independent.

Table 4: The correspondence rules between surface and

deep case.

Surface case that can be the left deep case Deep

case

Example of headline sentence

- wo-case

- no-case that is sahen-noun or adverbial form of

verb that cannot be time case

(O)

(1) shisetsu-wo sagasu (search the facility)

(2) touroku-chi-no sentaku

(choice of

the registered place)

- niyoru-case

- deno-case or no-case that cannot be (T) or (L)

- kara-case that cannot be source case

(I)

(3) koe-niyoru

sousa (operation by voice)

(4) jusho-de sagasu (search by the address)

(5) risuto-kara sagasu (search by the list)

- kara-case or karano-case whose depending

operation word is depended by (G)

(S)

(6) saisei-gamen-kara navi-gamen-ni modosu

(bring the replay screen back to the navigation

screen)

- ni-case or heno-case (G)

(7) genzai-chi-ni modosu

(bring back to the current place)

(8) audio-kinou-heno

setsuzoku

(connection to the audio function)

- deno-case or de-case with a status word (L)

(9) sukurouru-shita chiten-deno

sousa

(operation at the place that have been scrolled

)

- deno-case or no-case following a conditional

expression

(T)

(10) setsuzoku-ji-no kakunin

(confirmation at the time of connection)

(O): objective case, (I): instrumental case, (S): source case , (G):

goal case, (L): locative case, (T): time case

4.2.3 Different Type Case Merging

Different type case merging extracts case relations

by merging the different cases in the case relations

between the super and sub headlines when the both

verbs are the same, using rules shown in Table 5.

Table 5: The merge rules for different type cases.

Category The merge rules for different type cases

With

case element

if ( ec( Si ) = {verb/X, C1/A}

ec( Sj ) = {verb/X, C2/B}

headline-hiearchy( Si, Sj ) )

then ec( Sj ) = { verb/X, C1/A, C2/B }

Case element

ellipsis

if ( ec( Si ) = {verb/X, C1/A}

ec(Sj) = { verb/Y }

headline-hiearchy( Si, Sj ) )

then ec( j) = { verb/Y, C1/A }

ICSOFT 2008 - International Conference on Software and Data Technologies

304

For example, this processing can extract the case

relations “{verb/settei-suru (setting), object/touroku-

chi (registered place), instrument/kensaku-rireki

(search history)}” and “{verb/settei-suru (setting),

object/touroku-chi

(registered place),

instrument/jusho (address)}” form headline

hierarchy of the left column in Table 6.

Table 6: Case relations extracted from each headline.

Headline sentence Case relation

■ touroku-chi- no settei

(setting the registered

place)

- kensaku-rireki-

kara settei-suru

(setting from search

history)

- jusho-kara settei-suru

(setting from the address)

{verb/settei-suru (setting),

object/touroku-chi

(registered place)}

{verb/settei-suru (setting),

instrument/kensaku-rireki

(search history)}

{verb/settei-suru (setting),

instrument/jusho

(address)}

headline sentences

merge rules for different type case

(

with case elements

)

■ touroku-chi no settei

(setting the registered place)

kensaku-rireki kara settei-suru

(setting from the search history)

assignment of words

for variables

if ( ec( Si ) = { verb / X, C1 / A }

ec( Sj ) = { verb / X, C2 / B }

headline-hierarchy ( Si, Sj )

then ec( Sj ) = { verb / X, C1 / A, C2 / B }

instantiated rules

X = settei-suru (setting)

C1 = object

A = touroku-chi

(registered place)

C2 = instrument

B = kensaku-rireki

(

search histor

)

if ( ec( S1 ) = { verb / settei-suru, object / touroku-chi }

ec( S2 ) = { verb / setttei-suru,

instrument / kensaku-rireki }

headline-hierarchy (S1, S2 )

then ec( S2 ) = { verb / settei- suru, object / touroku-chi,

instrument / kensaku-rireki

}

case relations for each sentence

ec( S1 ) = { verb / settei-suru,

object / touroku-chi }

ec( S2 ) = { verb / settei-suru,

instrument / kensaku-rireki }

extracted case relations

ec( S2 ) = { verb / settei-suru,

object / touroku-chi,

instrument / kensaku-rireki }

Figure 2: Extraction of case relations by the merge rules

for different type cases.

The execution flow of the different type case

merge processing is shown in Figure. 2. Here, S1

denotes the headline in the upper layer (super

headline), S2 denotes the headline in the lower layer

(sub headline). This processing merges the deep case

“C1/A = object/touroku-chi (registered place)” in S1

with the deep case “C2/B = instrument/kensaku-

rireki (search history)” in S2, using the common

verb “X = settei-suru (setting)” in the case relations

extracted from each sentence. Then the processing

extracts the case relation {verb/settei-suru (setting),

object/touroku-chi (destination place),

instrument/kensaku-rireki (search history)} by

complement with the objective case omitted in S2.

4.2.4 Compound Word Case Analysis

Compound word case analysis extracts case relations

form compound words. By using the rules in Figure 3,

a new case relation {verb/Mn, case C/[M1, M2, …,

Mn-1]} is extracted, if the word ending is an

operation word Mn, and the noun [M1, M2, …, Mn-

1] is in the case frame dictionary as a case element or

a headline word.

if there is a compound noun N= [M1, M2, ..., Mn]

consists of the words M1,M2,...Mn, and

Mn is an operation word.

then

if there is a word Mn-1 as the ending word

of the term of case C in the case frame dictionary

then ec(N) = {verb/Mn, C/[M1, M2, ..., Mn-1]}

else if Mn-1is a sahen-noun

Mn-1 is an adverbial-form of verb

Mn-1 is a word stem of adjective/adjective-noun

then ec(N) = {verb/Mn, instrument/[M1, M2, ...,

Mn-1]}

else ec(N) = {verb/Mn, object/[M1, M2, ..., Mn-1]}

Figure 3: The case analysis rules for compound words.

Table 7: Examples of case relations extracted from

compound words in headlines.

Headline sentence

Extracted case relation

(1) sisutemu-onryo- no henkou

(modification of volume in the system)

(2) AV-onryo settei (AV-volume setting)

Ex.

{verb/henkou-suru (modify),

object/[sisutemu (system), onryou (volume)]}

{verb/settei-suru (setting), object/[AV (AV),

onryou (volume)]}

(1) ukai-ruuto-wo tsukuru(create a detour route)

(2) ruuto-henkou (change the route)

Ex.

{verb/tsukuru (create),

object/[ukai (detour), ruuto (route)] }

{verb/henkou (change), object/[rouuto (route)}

Table 7 shows examples of case relations

extracted from compound words. If the headline

sentence “

AV-onryo settei (AV-volume setting)” is a

compound noun N, N= [

AV (AV), onryo (volume),

settei (setting)], and the operating word is Mn =

“settei (setting)”. Then, Mn-1= “onryo (volume)” is

the ending word of the case “C = object” in the case

relation extracted from the headline sentence

“sisutemu-onryo-no henkou (modification of volume

in the system).” Therefore, a new case relation

TERM KNOWLEDGE ACQUISITION USING THE STRUCTURE OF HEADLINE SENTENCES FROM

INFORMATION EQUIPMENTS OPERATING MANUALS

305

{verb/settei-suru (setting), object/[AV (AV), onryou

(volume)]} can be extracted by the rules in Figure 3.

4.3 Extraction of IS-A Relations

4.3.1 Common Ending Word Processing

There is a conventional heuristics that an IS-A

relation holds when an inclusive relation as substring

between words that consists of compound words. This

heuristics incorrectly extracts IS-A(settei-naiyou

(setting content), naiyou (content))” whose IS-A

relation is not important for applications like the

operation help. Common ending word processing

extracts IS-A relations between terms by the rule

using the headline suffixes and the case frame

dictionary that are the specific expressions for the

headlines in the manuals.

Here, the headline words are the headline

sentences whose endings are the nouns depending on

“the headline suffixes” such as “gaiyou (overview)”,

“toha (what ~is)”, “oteire (maintenance)”, and

“sousa (operation)”, or the nouns without the

headline suffixes. For example, “onryou-settei-no

gaiyou (overview of volume setting)” is a headline

sentence, “onryou-settei (volume setting)” is a

headline word because “gaiyou (overview)” is a

headline suffix. We can make the dictionary of the

headline suffixes beforehand, because they are the

formulaic and domain independent.

if (there is a compound noun N= [M1, M2, ..., Mn]

consists of the words M1,M2,...Mn

there is a word Mn-1

as the ending word of the term of case C

in the case frame dictionary,

or the ending word of the headline word )

then IS-A(N, Mn)

Figure 4: The rule for the words with the common ending.

Using the rule in Figure 4, “IS-A(AV-onryo (AV-

volume), onryo (volume))” can be extracted from

“Sisutemu-onryo-wo henkou-suru (modify the

system volume)” and “AV-onryo (AV-volume).”

Because “{verb/ henkou-suru (modify),

object/[sisutemu (system), onryou (volume)]} ” is

extracted from sisutemu-onryo-wo henkou-suru

(modify the system volume)”, here, “compound

noun N=[AV (AV), onryo (volume)]”, “Mn= onryo

(volume)”, and “C=object”, then “IS-A(AV-onryo

(AV-volume), onryo (volume))” is extracted.

4.3.2 Same Type Case Refinement

Same type case refinement extracts the IS-A

relations using the case refinement rules for the

same type cases for nouns and for verbs in Table 8.

Table 8: The case refinement rules for the same type

cases.

Category The case refinement rules for the same type

cases

For

nouns

if ( ec( Si ) = {verb/X1, C1/A}

ec( Sj ) = {verb/X2, C2/B}

( X1 = X2 IS-A(X2, X1) )

headline-hiearchy( Si, Sj ) )

then IS-A(B, A)

For

verbs

if ( ec( Si ) = {verb/X1, C1/A1}

ec( Sj ) = {verb/X2, C2/A2}

( A1 = A2 IS-A(A2, A1) )

headline-hiearchy( Si, Sj ) )

then IS-A(Y, X)

For example, by using the case refinement rules

for the same type cases for nouns, IS-A(kisetsu-no

spotto (seasonal spot), osusume-jouhou

(recommended information)) and IS-A(hyaku-sen

(the hundred best), osusume-jouhou (recommended

information)) are extracted from headline sentences

“‘kisetsu-no supotto-wo sagasu (search the seasonal

spot) ’and ‘hyakusen-wo sagasu (search the hundred

best)’” which have super headline ‘osusume-jouhou-

wo sagasu (search the recommended information).’”

5 EVALUATION

5.1 Experimental Setting

We describe the evaluation result of our proposed

method using “an instruction manual for a car

navigation system (Mitsubishi 2005)” as input data.

The input data are 875 headline sentences with

hierarchical tags, and the output data are case

relations and IS-A relations. Here we use JUMAN

(Japanese Morphological Analysis System)

(Kurohashi, 2005a) and KNP (Kurohashi-Nagao

Parser) (Kurohashi, 2005b) for sentence analysis. In

evaluation, we calculate recall and precision by

comparing with the human-extracted relations as

answer relations. The numbers of answer relations

are 858 for case relations and 495 for IS-A relations.

ICSOFT 2008 - International Conference on Software and Data Technologies

306

5.2 Recall and Precision

We have compared that the proposed method using

headline hierarchy relation with the basic method

using one sentence analysis. Here we consider the

processing for the correspondence between surface

and deep case as the basic method for case

extraction, and the processing for the words with the

common ending as the basic method for IS-A

relation, because they are considered to be a

customization of the conventional methods to the

operating manual headlines. The evaluation result

for extracting case relations of the proposed method

has achieved 92.4% recall and 96.8% precision,

increasing 31.1% and 1.5% respectively from the

basic method as shown in Table 9.

The evaluation result for extracting IS-A

relations has achieved 93.9% recall and 89.9%

precision, increasing 21.0% and 1.0% respectively

from the basic method as shown in Table 10.

Table 9: An evaluation result of case relation extraction.

Basic Proposed Improvement

Recall 61.3% 92.4% 31.1%

Precision 95.3% 96.8% 1.5%

Table 10: An evaluation result of IS-A relation extraction.

Basic Proposed Improvement

Recall 72.9% 93.9% 21.0%

Precision 88.9% 89.9% 1.0%

The proposed method has almost achieved 90 %

of target recall and precision on the practical view.

The future issue is the establishment of the method

for suggesting related information like certainty

factors for human to correct false extraction.

5.3 Analysis of Extraction Failure

5.3.1 Extraction Leaks for Case Relations

The 22 relations (7.6% of all answers) of 65 total

leaks of case relations can be extracted by human

from one sentence as shown in Ex.1 and Ex.2 in

Table 11. They are difficult to extract because of the

lacks of clue in the target sentences. A future issue is

to establish a method of finding sentences, which

have the clue from the body of the manuals or Web

text, such as “denwa-de tsuuwa-suru (call by

telephone)”, “taju-de moji-wo housou-suru

(broadcast the characters by multiplex).”

The other 43 leak relations can be extracted by

human from headline hierarchy. The 30 leak

relations in them need knowledge about verbs

regarding the actions and the results of the operations

as Ex.3 in Table 12. The 13 leak relations in them

need to fill the ellipsis of case elements in the

sentences. To cope with former cases, it is required to

describe knowledge about basic verbs. To cope with

latter cases, it is good to consider the convention of

designing the headlines for the operating manuals.

Table 11: Examples of extraction leak for case relations.

Headline sentence

Extraction leak for case relation

denwa-no tsuwa (call by telephone) Ex.

{verb/tsuwa-suru (call),

instrument/dennwa telephone)}

moji-taju-housou (teletext broadcast) Ex.

{verb/housou-suru (broadcast),

object/moji (character),

instrument/tajuu (multiplex)}

torakku-risuto-wo hyouji-suru

(display the track list)

- arubamu-risuto-kara modoru

(return from the album list)

Ex.

{verb/modoru (return),

source/torakku-risuto (track list),

goal/arubamu-risuto (album list)}

adoresu-chou-wo kanri-suru

(manage the address book),

- touroku-suru (register)

Ex.

{verb/touroku-suru (register),

object/adoresu (address),

goal/adoresu-chou (address book)}

5.3.2 Extraction Errors for Case Relations

The extraction errors for case relations are 26 (7.6%

of all extracted relations). The 4 errors are caused by

the processing for the correspondence between

surface and deep case. For example, the case relation

{verb/tsuuwa-suru (call), object/denwa (telephone)}

is extracted as an error. The 13 cases are caused by

the case analysis processing for the compound word.

For example, the case relation {verb/housou-suru

(broadcast), object/moji-tajuu (multiplex)} is

extracted as an error of Ex.2 in Table 11. The other 9

errors are caused by the different type case merge

processing. For example, the case relation

{verb/tourokur-suru (register), object/adoresu-chou

(address book)} is extracted as an error of Ex.4 in

Table 11 The countermeasure of them are the same

as the extraction leaks.

5.3.3 Extraction Leaks for IS-A Relation

The 30 cases of extraction leaks for IS-A relations

(6.1% of all answers) do not meet the preconditions

TERM KNOWLEDGE ACQUISITION USING THE STRUCTURE OF HEADLINE SENTENCES FROM

INFORMATION EQUIPMENTS OPERATING MANUALS

307

of the same type case refinement processing as

shown in Table 12. It needs using the meaning of the

terms to extract them correctly. A future issue is a

method of suggesting how to loose the preconditions

of the extraction rules.

Table 12: Examples of extraction leak for IS-A relation.

Headline sentence

Extraction leak for IS-A relation

tsuushin-houhou-nitsuite (on communication method)

- i-moudo-setsuzoku (i-mode connection)

- internettto-setsuzoku (internet connection)

IS-A(i-moudo-setsuzoku (communication method),

tsuushin-houhou (i-mode connection))

IS-A(internetto-setsuzoku(internet connection),

tsuushin-houhou (internet connection))

5.3.4 Extraction Errors for IS-A Relations

The extraction errors for IS-A relations are 52

(10.1% of all extracted relations). The 45 errors of

them are caused by the processing for the words with

the common ending. For example, the IS-Arelation

IS-A(mokuteki-chi (destination place), chi (place)} is

extracted as an error from the headline sentences:

“mokuteki-chi-no kensaku (search of the destination

place)” and “touroku-chi-wo sagasu (search the

registered place)”. This is because the word chi

(place) that is the ending word of the compound

noun mokuteki-chi (destination place) is in the case

relation {verb/sagasu (search), object/ tourokuchi-

chi (registered place)}. The noun suffixes such as

chi (place), “mei (name)”, “chou (book)”, and “kata

(type)”, that appear in the word ending of the proper

nouns or technical terms are unnecessary for

extracting terms. It is preventable if we prepare the

dictionary for noun suffixes, however, we need to

create the noun suffix dictionary depending on each

application domain.

6 CONCLUSIONS

This paper describes a method for automatically

extracting case relations and IS-A relations from

operating manuals for help or search about

equipment manuals. The distinctive feature of the

proposed method is to use the hierarchical structure

of the headline sentences, while the conventional

methods are based on the processing of one

sentence.

We have confirmed the effectiveness of our method

by comparison of extraction accuracy of term

knowledge with the conventional methods.

REFERENCES

Waida, R., Hoshino, H., Tanaka, K., Shiratori K., 2007.

Navigation type Built-in manual on TV, Technical

Communication Simposium 2007. (in Japanese).

Kurohashi, S., Higasa, W., 2000. Dialogue Helpsystem

based on Flexible Matching of User Query with

Natural Language Knowledge Base, 1st SIGdial

Workshop on Discourse and Dialogue, pp. 141-149.

Kawahara T., Ito, R., Komatani, K., 2003. Spoken

Dialogue System for Queries on Appliance Manuals

using Hierarchical Confirmation Strategy, Proc. of 8th

European Conf. on Speech Communication and

Technology (Eurospeech2003), pp. 1701-1704.

Oishi, A., Matsumoto, Y., 1997. Detecting the

organization of semantic subclasses of Japanese verbs,

International journal of corpus linguistics, Vol.2,

No.1, pp. 65-89.

Kawahara, D., Kanji, N., Kurohashi, S, 2000. Japanese

Case Structure Analysis by Unsupervised Construction

of a Case Frame Dictionary, Proc. of the 18th

International Conference on Computational

Linguistics (COLING2000), pp.432-438.

Kurohashi, S., Nagao, M., Sato, S., Murakami, M., 1992.

A Method of Automatic Hypertext Construction from

an Encyclopedic Dictionary of a Specific Field, Proc.

of Third Conference on Applied Natural Language

Processing, pp. 239-240.

Kobayashi, N., Inui, K., Matsumoto, Y., Tateishi, K., and

Fukushima, F., 2004. Collecting Evaluative

Expressions for Opinion Extraction, Proc. of the First

International Joint Conference on Natural Language

Processing (IJCNLP-04), pp. 584-589.

Sudo, K., Sekine, S., Grishman, R., 2001. Automatic

Pattern Acquisition for Japanese Information

Extraction, Proc. of the First International Conference

on Human Language Technology Research, pp. 1-7.

Utsuro, T., Matsumoto, Y., 1998. Lexical Knowledge

Acquisition from Corpora, Proc. of JSPS-HITACHI

Workshop on New Challenges in Natural Language

Processing and its Application, pp.82-87.

Tsujii, J., Yamanashi, M., 1985. Cases and their

accreditation criteria, IPSJ SIG Notes, NL52-3, pp.1-8.

Information Processing Society of Japan. (in Japanese).

Fillmore, C. J., (translated into Japanese by Tanaka, H.,

Funaki. M.), 1975. Toward A Modern Theory Of Case

& Other Articles, Sanseido Bookstore (in Japanese).

Mitsubishi Electric Corporation., 2005. HDD Navigation

System CU-H9700 series Instruction Manual. (in

Japanese).

Kurohashi, S., Kawahara, D., 2005a. Japanese

Morphological Analysis System JUMAN 5.1. (in

Japanese).

Kurohashi, S., Kawahara, D., 2005b. KN parser

(Kurohashi-Nagao parser) 2.0 Users Manual. (in

Japanese).

ICSOFT 2008 - International Conference on Software and Data Technologies

308