=Paper=
{{Paper
|id=Vol-1312/jist2014pd_paper5
|storemode=property
|title=News Recommendation based on Semantic Relations between Events
|pdfUrl=https://ceur-ws.org/Vol-1312/jist2014pd_paper5.pdf
|volume=Vol-1312
|dblpUrl=https://dblp.org/rec/conf/jist/YokoKSTO14
}}
==News Recommendation based on Semantic Relations between Events==
https://ceur-ws.org/Vol-1312/jist2014pd_paper5.pdf
News Recommendation based on Semantic
Relations between Events
Ryohei Yoko, Takahiro Kawamura, Yuichi Sei, and Yasuyuki Tahara,
and Akihiko Ohsuga
Graduate School of Information Systems University of Electro-Communications,
1-5-1 Chofugaoka, Chofu-shi, Tokyo, 182-8585 Japan
{r-yokoh,kawamura,sei,tahara,ohsuga}@ohsuga.is.uec.ac.jp
Abstract. In recent years, “News Curation Services” that recommend
news articles on the internet to user have been popular. In this study,
we propose a new “News Curation Service” that collects and recom-
mends novel articles by using semantic relationships between events in
the news articles that a user feels interest. The semantic relationships be-
tween events are represented by Linked Data. In order to recommend the
news articles to the user, we create search queries by using the sentence
structure features. Finally, we collect the news articles on the internet
and recommend the articles to the user.
Keywords: Linked Data, News Recommendation System, Information
Retrieval
1 Introduction
Recently, web services such as “paper.li”1 and “The Tweeted Times”2 that au-
tomatically gather news articles, and recommend to users has been popular. The
users can easily get information with topicality and novelty by the web services.
Such web services are called “News Curation Services”. In this study, we pro-
pose a “News Curation Service” that recommends news articles with novelty
according to interest of a user. We focus on the semantic relationships between
terms of news articles that the user shows interest. The semantic relationships
between the terms are represented as Linked Data.
2 Related Work
Khrouf et al. [1] recommended event information to convert meta-information
(place, time, tag, genre, etc.) represented by Linked Data. They have built a
1
http://papper.li
2
http://tweetedtimes.com
Fig. 1. Summary of Our Approach
hybrid recommendation system by a method using Linked Data and collabo-
rative filtering. In other research, Ostuni et al.[2] has applied user’s implicit
feedbacks in the service, and recommended top n items by using Linked Data.
In recent years, many recommendation systems applying Linked Data have been
proposed. However, there is no recommendation system that converts phrases
and the semantic relationships within sentences to Linked Data. In addition, use
of Bag-of-Words model is a typical method in many recommendation systems,
but the model cannot represent the semantic relationships. Our approach can
incorporate the interest of the user by the semantic relationships in comparison
with the existing methods of using Bag-of-Words vectors.
3 Approach for Creating Linked Data
In this study, we recommend news articles which are collected on the internet
according to users’ preference. Fig.1 indicates a summary of our method. (1)(2)-
first, we collect news articles that a user feels interest and construct “User Pref-
erence Linked Data” by each sentence of the article. (3)(4)- next, we gather
articles that are candidates of the search queries on the internet to construct
“News Articles Linked Data”. Both Linked Data are composed of phrases and
semantic relations, that are extracted from sentences of the articles. (5)(6)- fi-
nally, if there is a similar subgraph between the two Linked Data, a search query
is created as the set of the words of the subgprah.
In order to construct the Linked Data, we extract sentence structures from
the articles. We define the sentence structures consists of Subject, Activity,
Object, Date, Time, and Location. Each phrase of a sentence is annotated
by one of these labels. We represent the sentence structures by using labeled
phrases. As an example of Fig. 1, the sentence structure is represented such
as “Kagawa→(Activity)→selected” and “selected→(Object)→Worst Eleven” in
the Linked Data. We use Conditional Random Fields(CRF) in a similar way as
the method of Nguyen et al.[3]. Nguyen et al. annotated the labels in web pages
and tweets in Twitter. We recreated a training data set and constructed a learn-
ing model for CRF from 100 news articles about soccer written in Japanese
in “Sponichi”3 . “Sponichi” is a famous internet news media in Japan. The
average F-measure of labeling with CRF by 10-folds cross-validations is Sub-
ject:63.19%, Activity:66.43%, Object:50.06%, Date:91.00%, Time:46.86%, and
Location:48.06%.
4 Evaluation
In order to create “News Article Linked Data”, we used 14,904 articles from
6/30/2013 to 7/1/2014 in “Sponichi”. Likewise, in order to create “User Prefen-
reces Linked Data”, we used 10 articles that a test user feels interest from 152
articles from 8/1/2014 to 8/5/2014 in “Sponichi”.
We asked to the test user whether the recommended article is fun(feels fun by
the information of the article), novelty(feels new) and serendipity(feels discovery)
to each testers. There are eight test users who are our university students. The
test users answered in 4 levels: “I thinks so”, “I think so a little”, “I don’t
think so a little”, and “I don’t think so”. We recommended top five articles that
are searched in Google by the search query created in the previous section. We
choose five search queries at random for each test user, so that we recommend
25 articles to a test user. We also defined that a score of the first article is 5
point, the second is 4 point · · ·, and the fifth is 1 point. Thus, the perfect score
is 15 for each query. In the case that the test users answers “I thinks so”, and
“I think so little” to the recommended article, the article was regarded as OK,
and we added the corresponding point.
We conducted comparison to the method using tf-idf method. We extracted
frequent words from 10 articles that the test user feels interest as our proposed
method. The search query is created by three of top 10 words calculated by
tf-idf method. Fig. 2 indicates the average scores by our method and the tf-idf
method. Our method was better than the user tf-idf method. Thus, our approach
is effective as a method of news recommendation. In particular, we confirmed
that it recommends news articles that are novel to the user.
3
http://sponichi.co.jp
Fig. 2. Comparative Evaluation Results
5 Conclusion and Future Work
In this paper, we proposed a “News Curation Service” by using semantic re-
lationships within sentences. Our approach incorporates interest of a user in
comparison with the existing methods. In the future, we will improve accuracy
of the labeling.
References
1. H. Khrouf, R. Troncy, Hybrid event recommendation using linked data and user di-
versity, Proceedings of the 7th ACM conference on Recommender systems, pp.185-
192, 2013.
2. Ostuni, Vito Claudio and Di Noia, Tommaso and Di Sciascio, Eugenio and Mirizzi,
Roberto: Top-N recommendations from implicit feedback leveraging linked open
data, Proceedings of the 7th ACM conference on Recommender systems, pp.85-92,
2013.
3. T.M. Nguyen, T. Kawamura, Y. Tahara, and A. Ohsuga: Self-supervised capturing
of users´ activities from weblogs, International Journal of Intelligent Information
and Database Systems, Vol.6, No.1, pp.61-76, 2012.
4. Lafferty, John D. and McCallum, Andrew and Pereira, Fernando C. N. : Condi-
tional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence
Data, ICML ’01 Proceedings of the Eighteenth International Conference on Ma-
chine Learning, pp.282-289, 2001.