Knowledge-based Identication of Emotional Status on Social Networks

Abstract

A knowledge based methodology is proposed for the content understanding and sentiment identication of the shared comments in social networks. The goal of this work is to retrieve the sentiment information associated to an opinion and classify it by its polarity and sentiment by means of a semantic analysis. Our approach implements knowledge graphs, similarity measures, graph theory algorithms and disambiguation processes. The results obtained were compared with data retrieved from Twitter and users' reviews in Amazon. We measured the eciency of our contribution with precision, recall and F-measure comparing it with the traditional method of just looking up concepts in sentiment dictionaries which usually assigns averages. Moreover an analysis was carried out in order to nd the best performance for the classication by using polarity, sentiment and a polarity-sentiment hybrid . A study is presented for remarking the advantage of using a disambiguation process in knowledge processing.

Keywords

• sentiment analysis
• knowledge engineering
• conceptual similarity

-

Nowadays the huge information transmitted on social networks has become a rich source of information for the human understanding as well as a way of expression where the users share their sentiment status and personal opinions through comments. The sentiment identication can classify comments as positive or negative(polarity) and unveil emotions such as anger, trust, sadness ,etc., on certain topics or users. Moreover the sentiments presented in the opinions can be relevant in the design of custom services, social plans for public health, marketing, e-commerce,etc.

On the hand sentiment analysis has become one of the fastest growing research areas in computer science due the outbreak of computer-based sentiment studies with the availability of subjective texts on the Web [ 16 ]. Furthermore the sentiment analysis has gained attention over the years in the general public as it is currently shown in Google trends [ 10 ].

Based on the previous motivation the present work aims in the identication of sentiment information in opinions on social networks. Our approach explores a content-based and semantic processing of the knowledge implicit in the comments. For each opinion we created a formal representation which it is associated with a sentiment and polarity. 2

Background

This section lists some relevant works related with the proposed methodology presenting their key features. As summary we present a discussion where we remark the main contributions of our work.

Describing briey some similar works related with sentiment analysis are: Anja Rudat[ 20 ] explored the criteria inuencing selection for retweeting in Twitter. Trying to discover relations on social networks Yuan Wang[ 24 ] proposed a methodology that inferred social relationships in microblogs based on physical interactions using user’s location records. The work of Garcia-Pablos [ 7 ] proposed an unsupervised system for the aspect-based sentiment analysis. One of the limitation of this work was the necessary to dene manually seed concepts and domains as input of the methodology. The work of Divya Sehgal et al., [ 21 ] proposed a real-time sentiment analysis using dictionaries but mostly focused on big data techniques that prioritize the velocity instead of a deeper analysis. Theodore Georgiou [ 8 ] proposed a community detection algorithm utilizing social characteristics and geographic locations.

Regarding the semantic processing the work of Shivam Srivastava [ 22 ] developed an algorithm to cluster places not only based on their locations but also their semantics in social networks, the contributions of this work was the geosocial clustering from check-in data. The work of Shuai Wang et al.[ 23 ] applied a semantics-based learning technique for a set of concepts previously labeled by grouping the target-related words in order to extract the semantics among words.

On the other side some researches related to social networks analysis are for instance the work of Shuiguang Deng[ 5 ] that proposed a recommendation service for the social networks with a trust enhancement method. Considering the inuence on social networks the work of Meng Jiang[ 14 ] studied the interpersonal inuence, the approach explains the importance of this factor for behavior prediction. Additionally Huang Liwei[ 12 ] explored the user preference, social and geographical inuence in order to recommend proper POIs (Point-of-interest). The machine learning implementation of Souvick Ghosh[ 9 ] processed the media text in order to determine the polarity and sentiment using manually labeled Facebook posts.

Reviewing the state-of-the-art, most of the researches worked with key social attributes that in general dismissed the semantics focusing in the lexical processing, keywords or explicit reactions in the social media. About the methodologies that implemented machine learning techniques they were based on a high quality large training datasets on a specic domain. On the hand our work handles the comments as excerpt of the knowledge, in this gap we prioritized the semantic level, sense and meaning of the whole comment. The proposal computed semantic similarity measures, conceptual expansion, graph theory algorithms and disambiguation using on a multi domain knowledge base. The methodology is exible which implies that the domains can be adjusted by just modifying knowledge base. 3

Methodology

This section describes the methodology in three main stages. The rst stage social networks discovery retrieves opinions from events or public proles by reading comments in photos, posts, videos, etc. The stage of knowledge processing constructs the formal representation for each comment. This module carries out processes of automatic knowledge graph construction enhanced by disambiguation. Finally the stage of sentiment analysis estimates the total polarity and main sentiment in the comments . 3.1

Social network discovery stage In the stage the comments are retrieved from public events or user proles on social network. This process obtains users, comments and the social graph’s structure. 3.2

Knowledge processing stage In this stage a content-based formal representation is constructed for each comment in the social network. This stage is composed by lexical preprocessing ,knowledge graph expansion, similarity measure and disambiguation. Lexical pre-processing. In the step the concepts in a comment are processed in order provide term matching with the knowledge base. The processes considered are: stop words elimination, tokenizer, stemming, and removal of unknown concepts in the knowledge graph.

Knowledge graph expansion. In this step the set of concepts obtained in the lexical processing are expanded on the knowledge graph until nding a common root for all their senses.

Let us dene G(C; R) as a knowledge graph with the set of concepts C and the set of relationships R; the knowledge base expansion (Ge)(equations 1, 2) for a concept c 2 C is the iterative process ( iteration) of discovering new concepts in knowledge graph (G) using semantic relations ( )(equation 4) that connect a origin concept c to the other destination concepts C (equation 3).

Ge0(c; G) = G0(C0; R0) = G0(fcg; ;)

Ge (c; G(C; R)) = G (C ; R ) C = ( = 0 fcg > 0 C 1 [ fy 2 C : x 2 C

1; (x; y) 2 Rg R = ( = 0 ; > 0 f (x; y) 2 R : x; y 2 C; x 2 C 1g (1) (2) (3) (4) Similarity Measurement. Once the concepts were expanded and an excerpt of knowledge was constructed from the previous stage, the next step is to establish similarity measures among all concepts. In order to accomplish this task two dierent approaches were implemented:

1) Automatically. It was implemented the similarity measure of conceptual distance DIS-C[ 19 ] that automatically establishes the similarity among concepts following the idea of visibility in the knowledge graph.

2) Manually. For each semantic relationship in the knowledge graph we established a weight in the range [ 0,1 ].

Disambiguation. In this stage a strongly connected graph GD(C; R) is created which is disambiguated and reduced (number of nodes and relationships) by a steiner tree algorithm. In the methodology we implemented the SketchLs algorithm[ 11 ] due the capability of handling large graphs. The disambiguation process starts counting the number of occurrences(senses)(Figure 1). If a concept has only one occurrence it implies that it has only one sense and it will participate in the disambiguation of the other concepts. On the other hand if a concept has more than one occurrence this concept has to be disambiguated.

During the disambiguation if the comment has only one concept and it has several senses then a dictionary of polysemy has to be consulted for nding most probable sense. On the other hand if the comment has more than a concept then the disambiguation will be computed.

Sentiment analysis stage Polarity calculation. In this step the polarity for comment is calculated P olarity(Comentx) taking into account the individual polarity of each concept P o (CP ). The process starts dividing the concepts in subsets Cx considering their positive or negative polarity Po( Cx)(see equations 5-6). In order to calculate the polarity Pot(Xg) for a set of concepts Xg the arithmetic mean is computed (equation 7). The total polarity of a comment P olarity(Comentx) is calculated by the sum of positive plus negative polarities XP and XN respectively(see equation 8). (5) (6) (7) (8) (9) (10) (11) XP = fCx j P o (Cx) > 0; Cx XP g XN = fCx j P o (Cx) < 0; Cx XN g

Pn P ot (Xg) = i=0 P o (Cx) ; Cx Xg

n P olarity(Commentx) = P ot (XP ) + P ot (XN ); XN ; XN Commentx Sentiment identication. In this step the sentiment status is identied in a comment Sentiment(Comentx) . For each concept Ci 2 Comentx, Ci it is expanded in the knowledge graph until nding one or more concepts linked to a sentiment Sx. The next process is to nd the the closest sentiment Sx to Ci by computing a shortest path algorithm and semantic similarities. Consecutively a pre-dened numerical weight W s(Cx) is assigned for the sentiment Sx which is located between the range [-1,-1] (equation 9). Once the weight of the sentiment was obtained the next step is to calculate the sentiment value Sen(C)x) for the concept Cx by multiplying the sentiment weight W s(Cx) by its polarity P o(Cx) (equation 10). Finally the sentiment status with the highest sentiment value Sen (Cx) is assigned to the comment Comentx (equation 11).

W s (Cx) = w (Sx) ; Cx ! Sx; w (Sx) 2 [ 1; 1]

Sen (Cx) = P o (Cx) W s (Cx)

Sentiment(Commentx) = max (fSen (Ci) j Ci 2 Comentxg)

The gure 2 presents the iterative process of expansion for nding the sentiment associated to a concept Cx in the knowledge base. When one or more concepts are located and they are linked to a sentiment then the Dijkstra algorithm with Fibonacci heap [ 6 ] is executed in order to select only one concept. This section presents the results after implementing the described methodology. It is divided in two subsections: knowledge bases and sentiment analysis. In this section we describe the knowledge base’s structure which is composed by: general knowledge graphs for common language understanding on several domains and sentiment dictionaries mapped into the knowledge graph.

General knowledge bases

WordNet[ 1 ] (version 3.1) is a large lexical database of English. Nouns, verbs, adjectives and adverbs are grouped into sets of cognitive synonyms (synsets). The Japanese WordNet[ 3,13 ] is similar to Wordnet for processing the Japanese language.

Open Multilingual Wordnet [ 4 ][ 3 ] provides access to wordNets in a variety of 34 languages merged into English WordNet.

Sentiment dictionaries

SentiWordnet [ 2 ] is a lexical resource that assigns polarity values to concepts in English WordNet.

NRC_emotion_lexicon [ 18,17 ] is a list of English words associated with eight basic emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust).

Sentiment Analysis In order to explain the results obtained in the sentiment analysis an example was processed from Twitter in the CNN News account. The comment considered is : a number of people feared dead after a dam bursts in kenya with hundreds left homeless ocials say. The table 1 presents the closest sentiment and a polarity value assigned by our methodology to each concept.

Id Wordnet-Concept

Sentiment with polarity

Finally the methodology estimates the total polarity and main sentiment presented in the comment(table 2).

Sentiment Polarity NRC_Anger -0.1875

Comment

Other relevant examples from the CNN news account are presented in table 3. We noticed a better classication using the basic sentiments instead of polarity. Sentiment

Polarity

Comment trust trust anger anger sadness sadness joy

Evaluation

This section measured the performance of our methodology comparing it with labeled data with sentimental information. We considered as a manual processing Twitter posts that we manually labeled and as automatic processing comments ranked by the users in amazon reviews. As traditional method (baseline) we proposed the process of only looking up concepts with polarity in dictionaries. 5.1

Sentiments evaluation on Amazon Reviews We evaluated our work with precision, recall and F-measure over 10 000 comments using the dataset Amazon reviews provided by the Stanford Network Analysis Project (SNAP)[ 15 ] and shared by Xiang Zhang [ 25 ]. In this dataset an user gives scores for products in the range of one to ve starts. We associated the scores with negative sentiments(anger,disgust, sadness,fear) and positive sentiments(joy, trust, anticipation, surprise) and a polarity value. The gure 3) presents the evaluation using polarity and sentiment with automatic and manual similarity measures during the semantic processing (polaritySemRelAuto, polaritySemRelManual, SSRelAuto and SSRelManual) and PolarityLexical(base line).

Additionally the gure 4 presents the evaluation with precision for the disambiguation process using polarity with automatic and manual similarity measures (polarityAuto, polarityManual). The results were compared to polarity lexical(baseline) with random sense selection (PolarityLexicalR1-R10). For this evaluation some comments were retrieved from Twitter and manually associated with a sentiment and polarity. The gure 5 presents the results only considering precision. The PrecisionLex (baseline) was calculated using only polarity. On the other hand PrecisionSS considered sentiment and computed a semantic analysis and a disambiguation process. In this experiment the PrecisionSS presented better results.

During the experiments we noticed that the methodology provides dierent results for specic sentiments (gure 6). For instance the sentiment anger or disgust performed better precision because usually the comments are more explicit.

On the other hand the joy was more complicated to identify because the usage of sarcasm or more implicit sentiments in the comments. In this paper a content-based methodology was proposed for the polarity calculation and sentiment status identication. The novelty of the presented work is the capability of handling the comments as excerpts of knowledge. We provided a mechanism of semantic processing using knowledge graphs, graph theory algorithms, semantic similarities and disambiguation. For the sentiment identication our work explored three dierent approaches (polarity, sentiment, sentimentpolarity hybrid) where the sentiment-polarity processing presented the best results.

We performed several experiments in order to compared our contribution with the traditional method of just looking up concepts in dictionaries(baseline) that usually counts polarity or concepts related with sentimental information and assigns averages.

Based on the experimental analysis the best relation precision and computing consumption was presented by the combination of sentiment, manual weights in semantic processing and disambiguation (SSRelManual). On the other the highest precision was obtained with automatic weights (SSRelAuto) costing a signicant increment in the usage of computing resources. Despite of the disambiguation presented a slightly better precision it provided the best combination of concepts for the construction of formal representations and thus better sentiment identication. The results obtained in the present work can be consulted at the github site: https://github.com/samscarlet/SBA.

Acknowledgments

This work was supported by CONACYT and JSPS KAKENHI Grant Number JP17H01789.

References

1. Princeton university "about wordnet. " wordnet . princeton university ( 2010 ), <http://wordnet.princeton.edu>

2. Baccianella , S. , Esuli , A. , Sebastiani , F. : Sentiwordnet 3.0: an enhanced lexical resource for sentiment analysis and opinion mining . In: LREC . vol. 10 , pp. 2200 2204 ( 2010 )

3. Bond , F. , Baldwin , T. , Fothergill , R. , Uchimoto , K. : Japanese semcor: A sensetagged corpus of japanese . In: Proceedings of the 6th Global WordNet Conference (GWC 2012 ). pp. 5663 ( 2012 )

4. Bond , F. , Foster , R.: Linking and extending an open multilingual wordnet . In: Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . vol. 1 , pp. 13521362 ( 2013 )

5. Deng , S. , Huang , L. , Xu , G. , Wu , X. , Wu , Z. : On deep learning for trust-aware recommendations in social networks . IEEE transactions on neural networks and learning systems 28(5) , 11641177 ( 2017 )

6. Fredman , M.L. , Tarjan , R.E. : Fibonacci heaps and their uses in improved network optimization algorithms . Journal of the ACM (JACM) 34(3) , 596615 ( 1987 )

7. Garca-Pablos , A. , Cuadros , M. , Rigau , G.: W2vlda: almost unsupervised system for aspect based sentiment analysis . Expert Systems with Applications 91 , 127137 ( 2018 )

8. Georgiou , T. , El Abbadi , A. , Yan , X. : Extracting topics with focused communities for social content recommendation . In: Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing ( 2017 )

9. Ghosh , S. , Ghosh , S. , Das , D. : Sentiment identication in code-mixed social media text . arXiv preprint arXiv:1707.01184 ( 2017 )

10. Google: Google trends .url: https://trends.google.com/trends/?geo=us.

11. Gubichev , A. , Neumann , T. : Fast approximation of steiner trees in large graphs . In: Proceedings of the 21st ACM international conference on Information and knowledge management . pp. 14971501 . ACM ( 2012 )

12. Huang , L. , Ma , Y. , Liu , Y. , Sangaiah , A.K. : Multi-modal bayesian embedding for point-of-interest recommendation on location-based cyber-physical-social networks . Future Generation Computer Systems ( 2017 )

13. Isahara , H. , Bond , F. , Uchimoto , K. , Utiyama , M. , Kanzaki , K. : Development of the japanese wordnet . ( 2008 )

14. Jiang , M. , Cui , P. , Wang , F. , Zhu , W. , Yang , S. : Scalable recommendation with social contextual information . IEEE Transactions on Knowledge and Data Engineering 26 ( 11 ), 27892802 ( 2014 )

15. Leskovec , J.: Snap: Stanford network analysis project ( 2014 )

16. Mntyl , M.V. , Graziotin , D. , Kuutila , M.: The evolution of sentiment analysis¿‰a review of research topics, venues, and top cited papers . Computer Science Review 27 , 1632 ( 2018 )

17. Mohammad , S.M. , Turney , P.D.: Emotions evoked by common words and phrases: Using mechanical turk to create an emotion lexicon . In: Proceedings of the NAACL HLT 2010 workshop on computational approaches to analysis and generation of emotion in text . pp. 2634 . Association for Computational Linguistics ( 2010 )

18. Mohammad , S.M. , Turney , P.D.: Crowdsourcing a wordemotion association lexicon . Computational Intelligence 29 ( 3 ), 436465 ( 2013 )

19. Rodguez Franco , H.: CÆlculo de la visibilidad de conceptos en ontologas . Ph.D. thesis , Instituto PolitØcnico Nacional. Centro de Investigacin en Computacin ( 2011 )

20. Rudat , A. , Buder , J.: Making retweeting social: The inuence of content and context information on sharing news in twitter . Computers in Human Behavior 46 , 7584 ( 2015 )

21. Sehgal , D. , Agarwal , A.K. : Real-time sentiment analysis of big data applications using twitter data with hadoop framework . In: Soft Computing: Theories and Applications , pp. 765772 . Springer ( 2018 )

22. Srivastava , S. , Pande , S. , Ranu , S. : Geo-social clustering of places from check-in data . In: Data Mining (ICDM) , 2015 IEEE International Conference on. pp. 985 990 . IEEE ( 2015 )

23. Wang , S. , Zhou , M. , Mazumder , S. , Liu , B. , Chang , Y. : Disentangling aspect and opinion words in target-based sentiment analysis using lifelong learning . arXiv preprint arXiv:1802 . 05818 ( 2018 )

24. Wang , Y. , Xiao , Y. , Ma , C. , Xiao , Z. : Improving users' demographic prediction via the videos they talk about . In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing . pp. 13591368 ( 2016 )

25. Zhang , X. , LeCun , Y.: Text understanding from scratch . arXiv preprint arXiv:1502.01710 ( 2015 )