Segmentation of Review Texts by using Thesaurus and Corpus-based

Word Similarity

Yoshimi Suzuki and Fumiyo Fukumoto

Interdisciplinary Graduate School of Medicine and Engineering, University of Yamanashi, Kofu, Japan

Keywords:

Text Segmentation, Reviews, Thesaurus, Wikipedia.

Abstract:

Recently, we can refer to user reviews in the shopping or hotel reservation sites. However, with the exponential

growth of information of the Internet, it is becoming increasingly difﬁcult for a user to read and understand all

the materials from a large-scale reviews that is potentially of interest. In this paper, we propose a method for

review texts segmentation by guest’s criteria, such as service, location and facilities. Our system ﬁrstly extracts

words which represent criteria from hotel review texts. We focused on topic markers such as “ha” in Japanese

to extract guest’s criteria. The extracted words are classiﬁed into classes with similar words. The classiﬁcation

is proceeded by using Japanese WordNet. Then, for each hotel, each text with all of the guest reviews is

segmented into word sequence by using criteria classes. Review text segmentation is difﬁcult because of short

text. We thus used Japanese WordNet, extracted similar word pairs, and indexes of Wikipedia. We performed

text segmentation of hotel review. The results showed the effectiveness of our method and indicated that it can

be used for review summarization by guest’s criteria.

1 INTRODUCTION

Recently, we can refer to user reviews in the shop-

ping or hotel reservation sites. Since a user’s cri-

terion is estimating the user review compared with

the information which a contractor offers, there is

a possibility that many information which is not

included in a contractor’s explanation is included.

These customer/guest reviews are often included var-

ious information about products/hotels which are dif-

ferent from commercial information provided by sell-

ers/hotel owners, as customer/guest have pointed out

with their own criteria, e.g., service may be very im-

portant to one guest such as business traveler whereas

another guest is more interested in good value for se-

lecting a hotel for his/her vacation. However, there

are at least seven problems as follows:

• There is a large amount of reviews for each prod-

uct/hotel.

• Each review text is short.

• There are overlapping contents.

• The wrong information may be described.

• It is not faithful to grammar.

• The terms are not uniﬁed.

• There are many miss spellings, words.

Moreover, it is difﬁcult to ﬁnd boundary for every

item. Because reviews explain one item by using two

or more sentences, or two or more items are explained

by using only one review sentence.

In this paper,we propose a method for reviewtexts

segmentation by using guest’s criteria, such as ser-

vice, location and facilities.

2 RELATED WORK

Text segmentation is one of the challenging tasks of

Natural Language Processing. It has been widely

studied and many techniques (Kozima, 1993) (Hearst,

1997) (Allan et al., 1998) have been proposed. Hearst

presented a method of TextTiling (Hearst, 1997),

which is based on lexical cohesion concerning to the

repetition of the same words in a document. Utiyama

and Isahara proposed a statistical method for domain-

independent text segmentation (Utiyama and Isahara,

2001). Hirao et al. proposed a method based on

lexical cohesion and word importance (Hirao et al.,

2000). They employed two different methods for text

segmentation. One is based on lexical cohesion con-

sidering co-occurrences of words, and another is base

on the changes of the importance of the each sentence

in a document.

381

Suzuki Y. and Fukumoto F..

Segmentation of Review Texts by using Thesaurus and Corpus-based Word Similarity.

DOI: 10.5220/0004140903810384

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2012), pages 381-384

ISBN: 978-989-8565-30-3

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

3 SYSTEM OVERVIEW

Figure 1 illustrates overview of our system. The sys-

tem consists of two modules, namely “Extraction of

criteria”, “Text segmentation”.

Figure 1: System overview.

4 EXTRACTION OF CRITERIA

Firstly, we deﬁne criteria as words which reviewers

notice in the reviews. In the text in Japanese, crite-

ria are represented followed by “ha”. For extracting

criteria in reviews, ﬁrstly we extract postpositional

particle “ha” from whole review texts. Next we ex-

tract words followedby “ha”, and ﬁnally, we collected

words which are index words of Japanese WordNet

from the extracted words as criteria words. Table 1

shows extracted criteria words with high frequency.

Table 1: Candidate words of criteria (top 5).

No words frequency

1 room 56,888

2 breakfast 25,068

3 meal 17,107

4 support 16,677

5 location 14,866

5 SIMILAR WORD PAIR

EXTRACTION

Review texts are written by many different people.

People may express the same thing by using differ-

ent expression. For example, “heya”, “oheya” and

“ruumu” are the same sense, i.e., room. Moreover,

two words such as “kyakushitsu”:(guest room) and

“heya”:(room) are often used in the same sense in the

hotel review domain while those are different senses.

We thus collected similar words from hotel re-

views by using Lin’s method (D.Lin, 1998).

Some technical terms do not frequently appear in

reviews even if we use large corpora. Therefore, we

applied smoothing technique to the hierarchical struc-

ture of semantic features. Firstly, we extracted simi-

lar word pairs using dependency relationships. De-

pendency relationship between two words is used for

extracting semantically similar word pairs. Lin pro-

posed “dependency triple” (D.Lin, 1998). A depen-

dency triple consists of two words: w, w

′

and the

grammatical relationship between them: r in the in-

put sentence. ||w, r, w

′

|| denotes the frequency count

of the dependency triple (w, r, w

′

). ||w, r, ∗|| denotes

the total occurrences of (w, r) relationships in the cor-

pus, where “∗” indicates wild card.

We used 3 sets of Japanese case particles as r. Set

A consists of 2 case particles: “ga” and “wo”. They

correspond subject and object, respectively. Set B

consists of 6 case particles. Set C consists of 17 case

particles. We selected word pairs which are extracted

by using two or three sets.

In order to extract the corresponding semantic fea-

ture of the new word, we extracted dependency triples

of the new word and the extracted words. Using some

extracted words, many types of dependencytriples are

extracted. For extracting the similar words from the

core thesaurus, I(w, r, w

′

) is calculated by using For-

mula (1).

I(w, r, w

′

)

= −log(P

MLE

(B)P

MLE

(A|B)P

MLE

(C|B))

−(−logP

MLE

(A, B, C))

= log

||w, r, w

′

|| × ||∗, r, ∗||

||w, r, ∗|| × ||∗, r, w

′

(1)

where P

MLE

refers to the maximum likelihood esti-

mation of a probability distribution.

Let T(w) be the set of pairs (r, w

′

) such

that log

||w,r,w

′

||×||∗,r,∗||

||w,r,∗||×||∗,r,w

′

is positive. The similarity

Sim(w

, w

) between two words: w

and w

is deﬁned

by Formula (2).

Sim(w

, w

)

∑

(r,w)∈T(w

)∩T(w

)

(I(w

, r, w) + I(w

, r, w))

∑

(r,w)∈T(w

)

I(w

, r, w) +

∑

(r,w)∈T(w

)

I(w

, r, w)

(2)

Table 2 shows the extracted pairs as similar word

pairs.

In Table 2, there are some notational variants.

In general, the pair of “morning newspaper” and

KEOD2012-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

382

Table 2: Results of extracting similar pairs using particle set

A, B, C.

No. word1 word2

1 favorable very favorable

(koukan) (taihen koukan)

2 route route

(michizyun) (ikikata)

3 stomach stomach

(onaka) (onaka hiragaga)

4 dust dust

(hokori) (hokori hiragana)

5 net Internet

(netto) (intaanetto)

“newspaper” and the pair of “breakfast voucher” and

“ticket” are not same meaning, however the two pairs

are mostly same meaning in hotel review texts.

6 SENTENCE PARTITIONING

AND TEXT SEGMENTATION

Compound sentences frequently appear in review

texts. Moreover, two or more criteria may be in-

cluded within a compound sentence. For example,

“The buffet-style breakfast is delicious, the room is

also large and the scent of the shampoo and rinse

in the bathroom are quite good”: “(chooshoku no

baikingu mo oishiidesushi, heyamo hiroishi, ichiban

kiniitteiruno ga heya ni oitearu shampuu to rinsu no

kaori ga totemo iito omoimasu)“.

It is necessary to divide one sentence into some

criteria. Fukushima proposed a method of sen-

tence division for text summarization for TV news

(Fukushima et al., 1999). They used rule based

method for sentence partitioning. In this paper, each

compound sentence was divided into some criteria

by using compound sentence markers and “cabocha”

((Kudo and Matsumoto, 2002)) which is a Japanese

dependency structure Analyzer.

Using results of sentence partitioning, we di-

vided text by criteria using lexical information of

Japanese WordNet, similarity of words and indexes

of Wikipedia. Figure 2 shows how to divide text.

Figure 2: Text segmentation method.

We decide border of criteria between 2 different

criterion words. The following algorithm is how to

decide border of criteria:

for(i=0; i<N; i++)

{

if(similarityA[i]<similarityB[i]) then

break;

}

border=i;

where N is number of sentences between two sen-

tences which have criterion words. similarityA[i] and

similarityB[i] are similarity between ith sentence and

“sentence A” and “sentence B”, respectively. Each

similarityA[i] is calculated by following algorithm:

similarityA[i]=0.0

for(j=0; j<M; j++)

{

similarityA[i]+=

cooccurrence(criterionA,wordj)

+cooccurrence(criterionA,simpair(wordj))/2

+cooccurrence(criterionA,hypernym(wordj))/2

+cooccurrence(criterionA,wikipedia(wordj))/2;

}

where M is number of words in sentence[i]. “cri-

terionA” is the criterion in “sentence A”. “wordj” is

jth word in sentence[i]. “simpair(wordj)“ is the word

of similar word pair of wordj. “hypernym(wordj)“

is the hypernym of wordj. “wikipedia(wordj)“ is the

most right noun of ﬁrst sentence of wordj. “occur-

rence(x,y)“ is the rate which x and y are co-occurring

in same sentence of the hotel review.

7 EXPERIMENTS AND

DISCUSSION

For the experiment, we used hotel review of Rakuten

Travel

. Table 3 shows Review data of the Rakuten

Travel.

Table 3: Reviews of Rakuten Travel.

amount of data 250MB

# of review text 350,000

# of hotel 15437

# of words for each review 375

# of reviews for each hotel 23

We used Japanese WordNet Version 1.1 (Bond

et al., 2009) as Japanese Thesaurus dictionary. Also

we used index words of Wikipedia

for dealing with

url= http://travel.rakuten.co.jp/ We used Rakuten travel

review data provided by Rakuten Institute of Technology

url=http://ja.wikipedia.org/ We used Wikipedia data of

2012-03-26.

SegmentationofReviewTextsbyusingThesaurusandCorpus-basedWordSimilarity

383

words such as named entities which are not index

words of Japanese WordNet. We employed Lin’s

method (D.Lin, 1998) for extracting similar word

pairs in hotel review texts.

We had experiments for dividing reviews into every

criterion. We used review texts of 5 hotels. The av-

erage number of review texts per hotel was 51.2. The

number of criteria consists of 256. Table 4 shows the

results of text segmentation.

Table 4: Results of text segmentation.

Our method TextTiling

Precision 863/1024=0.842 742/1024=0.725

Recall 863/902=0.925 742/1119=0.663

F-measure 0.882 0.692

We compared our method with TextTiling (Hearst,

1997) which is a well known text segmentation tech-

nique. TextTiling is a technique for subdividing texts

into multi-paragraph units that represent passages, or

subtopics. The discourse cues for identifying major

subtopic shifts are patterns of lexical co-occurrence

and distribution. TextTiling shows high perfor-

mance/accuracy for documents which have rather

long topics such as magazine articles. However, Text-

Tiling could not obtain good results as review text for

each criterion of is very short. As can be seen clearly

from Table 4, the results obtained by our method were

much better than those of TextTiling. This demon-

strates that lexical information such as similarity be-

tween words, hypernyms of words and named entities

were effective for text segmentation.

8 CONCLUSIONS

In this paper, we proposed a method for review texts

segmentation by guest’s criteria, such as service, lo-

cation and facilities. The results showed the effec-

tiveness of our method as the results attained at 0.84

precision, 0.92 recall, and 0.88 F-measure, and it

was outperformed the results obtained by TextTiling

which is used for text segmentation. Future work will

include: (i) applying the method to a large number of

guests reviews for quantitative evaluation, (ii) apply-

ing the method to other data such as grocery stores:

LeShop

, TaFeng

and movie data: MovieLens

evaluate the robustness of the method.

www.beshop.ch

aiia.iis.sinica.edu.tw/index.php?option=com docman&

task=cat view&gid=34&Itemid=41

http://www.grouplens.org/node/73

ACKNOWLEDGEMENTS

The authors would like to thank the referees for their

comments on the earlier version of this paper. This

work was partially supported by The Telecommuni-

cations Advancement Foundation.

REFERENCES

Allan, J., Carbonell, J., Doddington, G., Yamron, J., and

Yang, Y. (1998). Topic detection and tracking pilot

study ﬁnal report. In the DARPA Broadcast News

Transcription and Understanding Workshop.

Bond, F., Isahara, H., Uchimoto, K., Kuribayashi, T., and

Kanzaki, K. (2009). Enhancing the japanese wordnet.

In The 7th Workshop on Asian Language Resources,

in conjunction with ACL-IJCNLP.

D.Lin (1998). Automatic retrieval and clustering of similar

words. In Proceedings of 36th Annual Meeting of the

Association for Computational Linguistics and 17th

International Conference on Computational Linguis-

tics Proceedings of the Conference, pages 768–774.

Fukushima, T., Ehara, T., and Shirai, K. (1999). Partitioning

long sentences for text summarization. Journal of Nat-

ural Language Processing (in Japanese), 6(6):131–

147.

Hearst, M. A. (1997). Texttiling: Segmenting text into

multi-paragraph subtopic passages. In Association for

Computational Linguistics, pages 111–112.

Hirao, T., Kitauchi, A., and Kitani, T. (2000). Text seg-

mentation based on lexical cohesion and word im-

portance. Information Processing Society of Japan,

41(SIG3(TOD6)):24–36.

Kozima, H. (1993). Text segmentation based on similarity

between words. In Proceedings of the 31th Annual

Meeting, pages 286–288.

Kudo, T. and Matsumoto, Y. (2002). Japanese depen-

dency analysis using cascaded chunking. In CoNLL

2002:Proceedings of the 6th Conference on Natural

Language Learning 2002, pages 63–69.

Utiyama, M. and Isahara, H. (2001). A statistical model for

domain-independent text segmentation. In Proceed-

ings of the 39th Annual Meeting on Association for

Computational Linguistics, pages 499–506.

KEOD2012-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

384