37

Emotional News Recommender System

Ali Hakimi Parizi Dept. Computer Engineering

Razi University Kermanshah, Iran

Ali.h.13411@gmail.com

Mohammad Kazemifard Dept. Computer Engineering

Razi University Kermanshah, Iran

M.kazemifard@gmail.com

Abstract— With rapid advances of internet and overloading of information, it is important that we use some models and techniques to help users find proper data among massive flooding of information, especially in news domain that rapidly change. Recommender systems are a great help for achieving this goal. The current news recommender systems have focused on learning what users like to read based on their past activities and using methods for recommending news in a real-time manner, but none of them have considered emotion of news and how a user feels about an article in their recommendation process. Positive news can have a positive impact on user’s mood. In this work we aim to introduce a model for news recommender systems that can recommend news in a way to have a positive impact on the user’s mood. It utilizes both emotion of news and the user’s preference.

Keywords— News, Emotion, Content based recommender systems, Collaborative based recommender systems

I. INTRODUCTION Due to the information overloading and development of

internet in recent years because of advances in information technology, people face some difficulties to find their information among the huge amount of available information [1]. “Recommender Systems are software tools and techniques providing suggestions for items to be of use to a user” [2]. Recommender Systems have two basic components: products and users. These systems are becoming an essential tool in information retrieval and ecommerce. They work based on individual user’s interests, attributes that make them different from the usual information retrieval systems [3].

The goal of Recommender Systems is to direct users that are not capable of or do not have sufficient personal experience to find their favorable information from the pile of information offered on a website. For instance, in Amazon.com, the website uses a recommender system to help customers to find the books they are looking for. These systems aim to recognize items that may be suitable for user and for this they employ the user’s preferences and the item’s characteristics and also store the user’s feedback about the recommended items and use them in future to improve recommending process [2].

Recommender systems can be generally divided into three main approaches as follows [4]: Collaborative Filtering (CF), Content-Based (CB), and, Hybrid approach which combines the previous two methods. CF approach was used in the early researches that were done on Recommender Systems [5]. The system recommends items according to the users’ preferences. A characteristic of this method is that the more users rate items, the better will get the accuracy of recommendation. It works with taking into consideration the user’s past activities and also other users’ similar activities. In this way it recommends items that these users liked. There are two main paradigms in this approach: user-based and item-based techniques. In contrast to user-based CF, where it finds users that have similar past usage, the item-based CF technique finds items that are similar with the items that the user liked in the past, so it finds those items and then considers the users’ ratings for these items and recommends items that have higher ratings.

Another method in Recommender System is CB approach. It just uses the past activities of the user and characteristic of the items, recommending items that have conformity with the user’s preference, without involving data from other users. Since both of these systems have some shortcomings, researchers have decided to use Hybrid recommenders that combine two or more recommender’s techniques in different ways [2].

Fig. 1. Hybrid recommender that combines content based and collaborative based method

Over the last decade, researchers put their effort to improve efficiency of Recommender Systems. Nowadays, using context data in recommender systems have found high applications and researchers increase the efficiency of these systems by combining recommender systems with context-

Input CF based RS

Input CB based RS

Combiner Reco

2015 International Conference of Cognitive Science (ICCS), Institute for Cognitive Science studies, Iran

978-1-4673-7481-1/15/$31.00 ©2015 IEEE

38

aware systems [6]. Context has various definition but the most accurate one is “any information that can be used to characterize the situation of entities (i.e., whether a person, place or object) that are considered relevant to the interaction between a user and an application, including the user and the application themselves” [7]. One of the most important context data which plays an essential role in personalization of recommender systems is emotion [8].

In this paper we focused on a special usage of Recommender Systems. Today with the growth of technology and also internet, online news networks are getting more credit day to day and connecting to these networks and reading news from them are being more common [9]. However, as there are many articles available, finding interesting ones becomes more difficult. Until now no one has ever considered to sense affective information from news text with the aim to recommend news based on their emotions and effects that may have on users. So we intend to design a system to help users find information consistent with their preferences; hence, we design an emotional recommender system for news data. Today different characteristics of news recommender systems have been studied and also in some recent studies, researchers used context data to make them more personalized. Our goal is to use emotion for news and users’ preferences to recommend news.

The structure of the paper is defined as follows. In Section II we discuss some of the related works for news recommender systems and then in Section III we describe our model for extracting emotion from Persian texts. In Section IV we demonstrate recommendation approach that we choose for our news recommender and explain our case study and Section V concludes the paper.

II. RELATED WORKS In this section we briefly present some of the research

literature related to the news recommender systems and using context data in these systems.

The first recommender system that was originally designed by Group Lens for news systems was news Recommender System[8]. In this system CF method is implemented, recommending news to users while considering user’s historical activity and also ratings that other users gave to the news item. It finds users that have similarity in the pattern of past reading and then recommends news read by one of them to another user who didn’t see that news [9].

Because of some problems that CF has, researchers aim to using CB methods. Cold-start is one of the most important problems that CF systems have [10]. Whenever a new item is added to the system’s database, it does not have any rating, which makes it impossible to be recommended to any user, meaning that if a news item is not read by other users, the system can’t recommend it to any user [11],[12]. In CB, the system concerning user’s preference and news attribute and suggest news based on. But this approach has overspecialization problem [13]. The system only recommends highly similar items with the user’s previous preferences, and will never recommend other types of system’s items. Therefore, in news domain, the system does

not consider Hot news topics but just recommends routine and daily news that the user reads usually [14].

Because of such problems, researchers look for a way to solve them. So they use a combination of the two methods, i.e. a hybrid method. M. Balabanović and Y. Shoham [15] used two profiles for each user: one for long-term and another for short-term interests. Then by using CB, the system recommends news that are similar to the user’s historical reading and by employing CF, it finds hot topics and involves them in the recommending process.

More recently, several authors have been trying to use context in Recommender Systems. There is not much literature available about the combining news Recommender Systems and context aware systems. F. Garcin, K. Zhou, B. Faltings, and V. Schickel [16] used an AHP method to give news a rating by considering 4 attributes for news and then use Bayesian network to find out users’ interests. These attributes are: users’ profile, situation, rates and attributes. Users’ profiles contain their preferences and histories, while situation includes various types of context data, rates is the score that all users gave to that news and attributes are all characteristic of news stories. In the basic model, the authors considered a set of context data, but in implementation they just used location as a context data. Also in addition to use users’ past activity, S. Agarwal and A. Singhal [17], used the location pattern. They believe each person read different material in different places. For example a person reads economic stories in his office, sports event in cafe and entertainment news at home, so the system learns that what topics users like to read in each place and recommends news based on his location.

Gonzalez et al [8] is the first who used emotional context in Recommender Systems. They pointed out that, “emotions are crucial for user’s decision making in recommendation process. The users always transmit their decisions together with emotions.” Until now and with the current knowledge that we have, there is no work which uses emotion data in news recommenders. But automatic recognition of text emotion is getting more and more credit in these days [18]. For example sentiment system reads the news and tells that it has a negative, positive or natural sense [20]. Shaikh, Prendinger and Ishizuka [21] recently developed “SenseNet ” that is a linguistic tool to discover polarity values of words. In this approach with using WordNet, they assign a numerical value for some verbs, adjectives and adverbs manually. For others words, they detect their polarity values by ConceptNet. More recently, Mostafa et al [20] presented the Emotion Sensitive News Agent (ESNA).Its important part is SensNet and its purpose is to categorize news based on their emotion. In this work they used a cognitive model that defines 22 emotion states, but they just categorize news stories into eight emotion states.

In the next section we present a model to extract emotion of news from Persian text based on the user’s opinion and then use these information as a context data in recommending process in the way that makes a positive impact on the user.

39

III. EMOTION EXTRACTING

Here we aim to gather news from RSS and index them, and also extract main word of title and root of the verb with some processing on the title of news and use them for forecasting emotion of news. In the next step for recommending, we use users’ preferences, their past activities and also this emotion data in recommending equation. In this way news that have more positive emotion, get more score in recommending process.

Next, we described how we wanted to extract emotion of news. After reading previous works in this area, we applied methods used by G. Riva, F. Davide, W. A. I. (eds, and A. Boucouvalas, [23],vector space model and also a database that contains a set of sentences’ keywords and their emotion. At first we developed a program that collected title of news and saved them in a database. For this we used RSS and run our program to read RSS continuously and extract the title of news. Meanwhile when the title extracted, we gave it to a method that is responsible to extract sentence’s keyword from title and find the root of the verb. So far, we had a list of news and also a list of keywords that each set of them represented a sentence. Our purpose of these works is to build a dataset of keywords and their attached emotion. Then we developed an android program to show the news to users. When they saw and read the news we asked them about the emotion of news which would be attached to the set of keywords that gained from this title.

This was phase one of our project. In the second phase we wanted to implement an algorithm to recognize emotion of news from the title automatically. In the first step, like phase one, we extracted the keywords from the title. Then applied an algorithm to compute the weight of each. If we have D={d1,d2,….,dn} that is a set of all sentences and T={t1,t2,…..,tm} that is a set of all keywords, and also (w1j,w2j,…..,wtj) that Wij is weight of ti in dj, for measuring weight of each keyword we used tf-idf that tf is frequency of that keyword in that sentence and idf is inverse of frequency of that keyword in all sentences that we have. So the weight of ti in dj is as follows (Eq. 1):

Then we used normalized cosine to normalize weight that gains in this way, so for normalizing each of these wij, we used this equation (Eq. 2):

That wi is the weight of keyword i in the sentence. Next, we wanted to extract emotion of sentence. We tried to select the basic emotion so we used Ekman model that contains 6 emotions as follows: Happy, Sad, Fear, Anger, Hate and

surprise. Each emotion is considered as a class and sentences according to their emotions belong to one of these classes. To find out emotion of sentence, we used cosine similarity. At first we must obtain the vector of each class. If Mj is class of emotion j then we have (Eq. 3):

and then with cosine similarity (Eq. 4) we compute similarity between the sentence and our classes’ emotion to find out emotion of the sentence [24].

One of the problems in this project was considering a word and its derivatives equal. In a natural language, there are just limited words that can be considered as root words. In other words, each word has a root that has the same meaning. In Persian language, operation of derivation and creation of new words is done by combining root words and using different suffixes and prefixes. Operation of finding roots of words is done in two ways: a. using a table that contains words and root of each word that its name is table lookup method; b. using some algorithms that remove suffix and prefix to find root word and its name is affix remove. Because there is no source in Persian language to determine root of each word, we used affix removing method [25]. For this we used a simple root finder method. At first we saved a list of base words and roots of verbs in past and present. In Persian language, the structure of some sentences may be changed for example from “Subject+Object+Verb” to “Verb+Object+Subject”, making Persian a complex language for automatic NLP works. But in this work, we assumed that a sentence would be ended by verb. For each sentence, we first extract all words, then we look at the table for each word. If it exists, root finder operation won’t be executed but if it doesn’t exist in the table, we do affix remove method as much as possible. This process goes on until no other letter can be removed. Also in each cycle that the word matches to a word in the table, the process will stop.

Another case that we must pay attention to is negative verbs in the sentence. In most sentences with a negative verb, emotion of sentence is the opposite of emotion of sentence without negative verb. Therefore in this research we check if the verb is negative, emotion of sentence with positive verb is obtained first and then the opposite emotion is put for it.

IV. RECOMMENDING MODEL Until now we can extract emotion of news with

considering user’s opinion. In the next step we want to describe the recommendation part. As we said, there are various techniques to build a Recommender System. Because of the disadvantages that CB and CF have, we decide to use a hybrid method, so we utilize two recommender system

40

techniques for modeling user preference and past activities. In this case, we model short term and long term preferences of the user using a CB recommender system. Short term preferences are more dynamic and show interests that change constantly, but long term preferences are more stable and show interests that don’t change very much with the passage of time. For example a person during World Cup events prefers to read news about World Cup, but he/she always follows up political news. To learn about user’s interests, first we must find out about his/her preferences. We can get users’ preferences in implicit or explicit ways. In explicit way, we can get preferences with questionnaires and in implicit way, we can log the user’s activity and click behavior. In addition, we use a collaborative recommender system to recognize hot news topics and involve them in the recommendation process. Hot topics are recognized by analyzing all users’ preferences. News is in this list if it was read by many users. By using a collaborative model, our system can recommend some news that the user might have not read like them before, so in this way our system can achieve diversity in its offers.

Finally we must combine emotion data in recommendation process. The system recommends news in a sequence, according to emotions of news obtained from the model that we described in previous section, to cause a positive impact on user’s mood.

Fig. 2. Combining Emotion data in recommendation process

V. RESULT We are completing implementation as well as test

scenarios. After our dataset is complete, we aim to test the accuracy of our method for extracting emotion of news from title. In the next stage, our plan is to implement two news recommender systems, one using a classic recommender technique that learns user’s preference and the other using this model to discover emotion of news with respect to interests of users.

In the test phase we aim to test our system with 8 persons (4 men and 4 women), and we don’t tell them which one of these systems they use. For evaluating we use some questionnaires that users can reflect their opinions about the systems in them. Based on these questionnaires and also on using some data about how much users click on recommended news, we compare these two models.

VI. CONCLUSION In this paper we introduced an emotional news

recommender system that is able to recommend news with considering a user’s preference and emotion of news in the way that have a positive impact on the user’s mood. Due to exponentially growth of internet and popularity of news networks, with using this model in news networks we can have a positive impact on user’s mental health.

First of all we designed an application to gather a dataset for our system. These data include the keywords of sentences and an emotion that represents emotion of sentence. We collected these data from users.

In our system we use a vector space method to extract emotion of sentence with using the dataset that we gather in the previous phase. Our future work is to implement news recommender system and use emotion data as context data to improve the efficiency of our system and also make a positive impact on user’s mood implicitly.

References

[1] H. Costa and L. Macedo, “Emotion-Based Recommender System for Overcoming the Problem of Information Overload,” in Highlights on Practical Applications of Agents and Multi-Agent Systems SE - 18, vol. 365, J. Corchado, J. Bajo, J. Kozlak, P. Pawlewski, J. Molina, V. Julian, R. Silveira, R. Unland, and S. Giroux, Eds. Springer Berlin Heidelberg, 2013, pp. 178–189.

[2] F. Ricci, L. Rokach, and B. Shapira, “Introduction to Recommender Systems Handbook,” in Recommender Systems Handbook SE - 1, F. Ricci, L. Rokach, B. Shapira, and P. B. Kantor, Eds. Springer US, 2011, pp. 1–35.

[3] R. Burke, “Hybrid Recommender Systems: Survey and Experiments,” User Model. User-adapt. Interact., vol. 12, no. 4, pp. 331–370, Nov. 2002.

[4] G. Adomavicius and A. Tuzhilin, “Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions,” Knowl. Data Eng. IEEE Trans., vol. 17, no. 6, pp. 734–749, Jun. 2005.

[5] D. Billsus and M. J. Pazzani, “Learning Collaborative Information Filters,” in Proceedings of the Fifteenth International Conference on Machine Learning, 1998, pp. 46–54.

[6] Z. Yujie and W. Licai, “Some Challenges for Context-aware Recommender Systems,” in 5th International Conference on Computer Science and Education (ICCSE), 2010, pp. 362–365.

[7] M. Baldauf, S. Dustdar, and F. Rosenberg, “A Survey on Context Aware Systems,” Int. J. Ad Hoc Ubiquitous Comput., vol. 2, no. 4, pp. 263–277, Jun. 2007.

[8] Y. Zheng, R. D. Burke, and B. Mobasher, “The Role of Emotions in Context-aware Recommendation.,” in Decisions@RecSys, 2013, vol. 1050, pp. 21–28.

[9] C. Lin, R. Xie, X. Guan, L. Li, and T. Li, “Personalized news recommendation via implicit social experts,” Inf. Sci. (Ny)., vol. 254, pp. 1–18, Jan. 2014.

Item database

User database

Matching user profile and items profiles

Recommendation

Emotion Extraction

41

[10] F. Garcin, B. Faltings, O. Donatsch, A. Alazzawi, C. Bruttin, and A. Huber, “Offline and online evaluation of news recommender systems at swissinfo.ch,” in Proceedings of the 8th ACM Conference on Recommender systems - RecSys ’14, 2014, pp. 169–176.

[11] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. Riedl, “GroupLens: An Open Architecture for Collaborative Filtering of Netnews,” in Proceedings of the ACM Conference on Computer Supported Cooperative Work, 1994, vol. pp, no. 1, pp. 175–186.

[12] A. I. Schein, A. Popescul, L. H. Ungar, and D. M. Pennock, “Methods and Metrics for Cold-start Recommendations,” in Proceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 2002, pp. 253–260.

[13] J. Liu, P. Dolan, and E. R. Pedersen, “Personalized News Recommendation Based on Click Behavior,” in Proceedings of the 15th International Conference on Intelligent User Interfaces, 2010, pp. 31–40.

[14] D. M. Desyaputri, A. Erwin, M. Galinium, and D. Nugrahadi, “News recommendation in Indonesian language based on user click behavior,” in International Conference on Information Technology and Electrical Engineering (ICITEE), 2013, pp. 164–169.

[15] M. Balabanović and Y. Shoham, “Fab: Content-based, Collaborative Recommendation,” Commun. ACM, vol. 40, no. 3, pp. 66–72, Mar. 1997.

[16] F. Garcin, K. Zhou, B. Faltings, and V. Schickel, “Personalized News Recommendation Based on Collaborative Filtering,” in Proceedings of the The IEEE/WIC/ACM International Joint Conferences on Web Intelligence and Intelligent Agent Technology - Volume 01, 2012, pp. 437–441.

[17] S. Agarwal and A. Singhal, “Handling Skewed Results In News Recommendations by Focused Analysis of Semantic User Profiles,”

in International Conference on Optimization, Reliabilty, and Information Technology (ICROIT), 2014, pp. 74–79.

[18] K. F. Yeung, Y. Yang, and D. Ndzi, “Context-Aware News Recommender in Mobile Hybrid P2P Network,” in Second International Conference on Computational Intelligence, Communication Systems and Networks (CICSyN), 2010, pp. 54–59.

[19] Y. Noh, Y. Oh, and S.-B. Park, “A location-based personalized news recommendation,” in International Conference on Big Data and Smart Computing (BIGCOMP), 2014, pp. 99–104.

[20] C. Strapparava and R. Mihalcea, “Learning to identify emotions in text,” in Proceedings of the ACM symposium on Applied computing - SAC ’08, 2008, pp. 1556–1560.

[21] S. Mostafa Al Masum, H. Prendinger, M. Ishizuka, and S. M. Al Masum, “Emotion Sensitive News Agent: An Approach Towards User Centric Emotion Sensing from the News,” in IEEE/WIC/ACM International Conference on Web Intelligence (WI’07), 2007, pp. 614–620.

[22] S. Mostafa, A. Masum, H. Prendinger, and I. Mitsuru, “SenseNet : A Linguistic Tool to Visualize Numerical-Valance Based Sentiment of Textual Data,” 2004.

[23] G. Riva, F. Davide, W. A. I. (eds, and A. Boucouvalas, “Real Time Text-to-Emotion Engine for Expressive Internet Communications.” 2003.

[24] G. Salton, Automatic Text Processing: The Transformation, Analysis, and Retrieval of Information by Computer. Boston, MA, USA: Addison-Wesley Longman Publishing Co., Inc., 1989.

[25] M. Tashakori and M. R. Meybodi, “Designing an Automatic Root Finder for Farsi,” in Proceedings of The Seventh Annual International CSI Computer Conference, 2002, pp. 402–409.