Personal Life Event Detection from Social Media Smitashree Choudhury Harith Alani Knowledge Media Institute Knowledge Media Institute The Open University The Open University United Kingdom United Kingdom smitashree.choudhury@open.ac.uk h.alani@open.ac.uk ABSTRACT tool, nonetheless most popular online systems are carrying Creating video clips out of personal content from social me- huge amount of data created by individual users in the form dia is on the rise. MuseumOfMe, Facebook Lookback, and of texts, videos, and photos. While technology for data cre- Google Awesome are some popular examples. One core chal- ation and storage has significantly matured and efficiently lenge to the creation of such life summaries is the iden- managed, accessing, managing and processing of such data tification of personal events, and their time frame. Such is still a challenge and can be done by fews experts. Due to videos can greatly benefit from automatically distinguishing the lack of efficient data access mechanism available to nor- between social media content that is about someone’s own mal users, most of the historical data tend to be forgotten wedding from that week, to an old wedding, or to that of a or will remain unused. friend. In this paper, we describe our approach for identi- fying a number of common personal life events from social Access and reuse of such information trove will provide greater media content (in this paper we have used Twitter for our insight about the individual user, their preferences, and sit- test), using multiple feature-based classifiers. Results show uational dynamics and result in many useful applications that combination of linguistic and social interaction features e.g. personalised healthcare, customised training and edu- increases overall classification accuracy of most of the events cation, social and community engagement application and while some events are relatively more difficult than others life stories. To this end, mining and analysing such con- (e.g. new born with mean precision of .6 from all three mod- tent could help identifying one’s life milestones and salient els). events. Identifying interesting and important moments in one’s timeline on social media is valuable to services such as Keywords Facebook Lookback and Google Awesome, which generates Social Web, social media, event detection, personal life events short video clips for users to summarise and visualise their timelines. 1. INTRODUCTION In realisation of the importance of events on social media, With the wide spread of social media sites (e.g. Twitter, Facebook 2 has recently generated millions of 1 minute look- Facebook, YouTube), millions of of people use them on daily back videos of content from users’ timelines. Over 270 mil- basis to communicate and share information on a wide vari- lion video rendered and over 200 million users watched their ety of events, ranging from world events (e.g. World Cup), look back movie in the first two days and more than 50% to personal events (e.g., Wedding, Graduation). Use of these shared their movie. A project like Intel’s Museum of Me3 systems serves the multitude of purposes of knowledge shar- follows a similar line to collect data from user’s Facebook ing, information communication, event organisation, profes- profile and generate a short video. Purpose of our work (per- sional collaboration, political expression, as well as social- sonal life event detection) is a sub-objective of the broader isation. To put in perspective, more than 500 million of research objective in similar direction i,e, automatic creation tweets generated in a day1 , millions of photos are uploaded of digital documentaries from social media content including to Facebook every day. There may be differences in terms interesting and relevant life moments and events. of content volume created on different platforms depending on the personal preferences and the perceived purpose of the Event detection from social media content has so far been fo- 1 https://blog.twitter.com/2013/new-tweets-per-second- cused on detecting world events such as earthquakes [Chile, record-and-how japan], political protests, elections (US, Germany, UK ) and planned public events such as entertainment award func- tions (Oscar, Golden Globe), academic events (conferences), sports event (Olympic). However, detection of personal life events have been mostly overlooked, and only mildly inves- tigated for content recommendation [cite]. Objective of this piece of is to automatically identify interesting and impor- 2 https://code.facebook.com/posts/236248456565933/looking- back-on-look-back-videos 3 http://www.intel.com/museumofme/r/index.htm tant life events of individual users from their social media resurgence of interest in detecting social topics and events content, which can be part of their personal digital story- in this new domain[?]. We have been motivated by the need book or memory archive. In this work, we have taken Twit- to identify personal life events, which have a great personal ter as the test platform and will extend our research to other value when aggregated over time and location. One of the systems such as Facebook, Instagram, Pininterest in our fu- prerequisites of such a system is the identification of content ture work. reporting a real event. Events can be planned events such as cultural events, tech conferences, music award functions, Detecting personal events is non-trivial and may require a elections or sports event or unplanned events for example, combination of multiple approaches for a robust detection natural disasters, earthquack [?] and even generic events result. Unlike public events or events concerning celebrities such as breaking news events are subject of few studies [?][?]. and well-known personalities, personal events may not be Existing studies cover both planned and unplanned events characterised by high activity volume and additional sources with varying degrees using both machine learning and text of information e.g. blogs or Wikipedia. These events are analysis techniques. Benson et.al.[?] reported detecting con- limited to the concerned person and to her immediate social cert events from social media stream using city calendar as network (friends and family). In addition to the above prob- a target list. Agarwal et. al.[?] detected events such as lems, microblog sites like Twitter bring its own complexi- factory fire, labor strike from Twitter stream using a com- ties with short, informal and noisy content. Any meaning- bination of local sensitive hashing and location dictionary. making task on these content has to deal with these idiosyn- Weng and Lee[?] proposed event detection with cluster- crasies. Next, we will briefly delve into the concept of a per- ing of word bursts from tweets. Authors in [?] proposed a sonal event before going into the details of the experimental natural disaster alert system using Twitter users as virtual work. sensors. In their work, they were able to calculate the epi- centre of an earthquake by analyzing the delays of the first 1.1 Personal Life Events messages reporting the shock. Social media centric event Personal life events range from recurring events such as birth- detection also covers non textual data such as photos and days and anniversaries, to very occasional and uncommon videos, Chen et al.[?] discovered social event from Flickr events, such as work promotions, and relocation. Events can photos by using both user tags and other metadata including also be further categorised on an affective scale, from highly time and location (latitude and longintude). Firan et.al[?] positive and pleasant events to to unpleasant events, such as explored tags, title and description to classify pictures into illnesses or accidents and deaths of loved ones. In this pa- event categories. Some of the popular approaches used for per, we focus on 5 life events (4 positive and 1 negative) i.e. event detection are spatio-temporal segmentation[?], burst graduation, marriage/engagement, new job, birth of child, analysis in word signals, clustering as well as topic detection and surgery. Our motivation to start with these events in- techniques. spired by a study [?] which lists 6 important memorable life events are ”Beginning school”, ”first full time job”, ”Falling To the best of our knowledge, we found no prior studies on in love”, ”Marriage”, ”Having children;”, ”Parent’s death”. personal life event detection from social media except one reported in [?] where authors tried to detect two life events The main contributions of this paper are not on algorithm ”marriage” and ”employment” and bears some similarity to and its efficiency, but rather on presenting evidence that our work. Our focus is on user level event detection that can with effective combination of existing methods and social be used to build individual digital storyboards form histor- media data, we can analyse and detects important and criti- ical data. cal moments of individuals life., hence the contributions are: • a thorough study of five personal life events and their 3. PERSONAL EVENTS ON TWITTER We now define the concept of personal life event in the con- idiosyncrasies as reported in social media especially in text of Twitter message stream and provide a definition of Twitter . the problem that we address in this work. • detection of life events using both content and inter- action features. Definition of term ”event” differs from domain to domain ranging from Philosophy to cognitive psychology to com- puting. Despite a lack of uniform definition of the term it This paper is organised as follows: In section 2 we review re- embeds a few generic characterstics such as time, partici- lated work in the field of event detection in social media and pating objects and a location. In this context, we define in section three, we briefly describe how personal life events an event as a real world occurrence with an associated time are reported on twitter and their characterisation. Section 4 period and one or more participating objects/agents at a describes our approach which includes feature selection a nd certain location which may or may not be explicitly appar- model construction followed by discussion and conclusion in ent in tweet messages. According to this definition a tweet section 5. needs to reflect a time interval when the event has occurred involving either the user or someone connecting to the user 2. RELATED WORK as the participating agent. Based on this abstract notion, Event detection is now a new research subject, and has been we looked into the real data to confirm or re-arrange the def- part of studies on topic detection in news stories and other inition and devise a strategy for detecting personal events. text documents [?]. Social media bought multi modal con- tent created by both professional and amateurs leading to a 3.1 Dataset The second dimension where the event reporting differs is on As a first step, we collected tweets using Twitter streaming participating agent or affected subject. Event tweets are ei- API4 which allows to crawl some portion of public tweets as ther about the user who created the tweet or about someone and when it comes. We restricted tweets to English language else known to the user and in some cases, about an undefined only and crawled for 3-4 hours per day for three weeks. The group of people e.g. group of students. Since our focus is on entire dataset contained around 4 million tweets. Ratio of personal events, ideally we should target self-reported tweets event tweets to non-event tweets is expected to be extremely and ignore the rest. But resolving an event to a participat- skewed as the targeted events are very specific and user cen- ing agent needs advanced semantic role labelling which will tric. So the next logical step is to use a filter mechanism to be our next step of this ongoing work. For this paper, we segregate the event related tweets from the rest and process restricted our attention to generic event detection, hence in- further. For this initial segregation, we extended the event cluded all the tweets irrespective of who the affected subject query with synonyms and related terms and phrases (shown is. in Table 1). These related terms are mainly synonyms and terms commonly known and used to describe the event of Based on this generic definition, we proceed with our actual interest. Use of related terms with the main event terms experiment task that starts with feature extraction. were intended to widen the coverage where users might not be using the exact terms to describe the main events. After 4. FEATURE EXTRACTION filtering we got 9168 tweets for marriage event, 2570 tweets After filtering event related tweets from the non-event tweets, for graduation, 3192 tweets for surgery, 3661 for new job and we extracted different types of features [?] to be used for 2954 tweets for new born. A question may arise about those building event classifiers. We examined several feature cate- tweets where the event term may be absent yet the implicit gories describing different aspects of tweets and users. Specif- semantics reflects a real event for example. ”Welcome to the ically we considered lexical, sentimental and social interac- new member of our family”. However, we agree such kind tion features. of possible omissions with the present approach and intend to capture them with contextual and historical information as part of our future work. The resulting filtered datasets 4.1 Textual Features Event term: The basic lexical feature of an event is the still contain many irrelevant tweets. For example, ”family event term itself and most closely related terms or have brought a 2nd lawsuit against her, this time to try to its synonym ”#graduation, convocation” for the event annul her marriage” is not about a marriage event though it graduation. The synonyms are extracted from Word- contains the keyword. Our task is identify such tweets from net5 genuine event tweets by means of binary classification. Co-occurring textual Features are the features of a term that co-occur significantly along with the event term Table 1: Events are their related words. for example, ”cap”, ”dress”, ”present”, ”prom”,”party” Event terms Related Terms are some of the frequently occurred terms for gradu- Marriage ”Wedding”,”Tied the knot”,”married” ation, while ”prayer”, ”hospital” for surgery. Presence Graduation ”Convocation”,”commencement ” of these terms along with the main event term is ex- New Job ” new position”,”first day at work”,”job offer” pected to boost the detection process. Co-occuring New Born ”Baby boy”,”baby girl”, ”new born” terms were extracted from various tag based social me- Surgery ”Operation” dia sites such as Flickr, instagram where terms are de- scribed with highly related terms. are These features are event specific and treated as binary values i.e. 1 Manual inspection of these tweets revealed that event re- for presence otherwise 0. porting tends to happen at three time spans; part, present, and future. We also noticed three categories of participating Temporal terms: This feature reflects the presence of time agents (self, others individual and general public). Examples terms in a tweet. Since the content are about an event, of such diversities are shown in table 2. it is intuitive to assume that some reference to time is natural and required by definition. For this feature, In light of these findings, defining a personal event seems to we used LIWC’s time category which includes 68 time be more tricky and imprecise. Two pertinent questions here terms. are how to resolve the time reference associated with the Person reference terms: Since these events are about per- event and how to associate the right subject (participating sonal life event one or more reference terms reflect- agent) with the event. In this study we are only focusing on ing social relation is expected when the event is about the events where the time reference can be resolved to a spe- somebody other than the poster, or self reference if the cific time point within a month time interval by automatic event is about the user. means. One such example is ”I graduated yesterday”, ” 26 days to graduation”. In both cases, the time of the event Sentiment: personal events are expressed with rich emo- can be resolved with help from the timestamp attached to tions both for pleasant or unpleasant events. Senti- the message. However, ambiguous time references such as ments are detected by Sentistrength [?] library and ”graduation is so close yet so far”, ”marriage in few weeks proved to be good for social media sentiment detec- time” are ignored. tion. Value of this feature ranges from -5(negative) to +5(positive) while +1 to -1 considered as neutral. 4 5 https://dev.twitter.com/docs/api/streaming http://wordnet.princeton.edu/ Table 2: Events and their examples from Twitter. Event Examples Marriage Kansas City here we come! It’s happening! My sister’s marriage this weekend!! :) 8 years ago this day , married to the most loving man on this earth. Congratulations to my beautiful friend, @SheridanMillls, who tied the knot today! ??? Graduation Happy graduation day, bebe! Congrats cutie pie! http://t.co/YqgNgK9WMw Graduation is just around the corner. Time to start planning programs and certificates. Talk to our print consultants today! 3 sets of graduation picture next week! Hahaha. At last! :) New Job First day of a new job.... Kind of dreading it. #officeassistant Starting my new position today. Ayy lmao. Shout out to my cuz Quincy Johnson aka Q. On his new Executive Chef position! ??? New Born My baby girl is here! Introducing: Halen born naturally May 3rd @ 4:43 pm. Exactly 3 weeks till my babyshower & almost 7 weeks till my baby boy Is born ? Surgery Good luck on your surgery today @chloebieber ear surgery ??it went well Everyone please continue to pray for Karlie these next 5 hours. She just went back for her brain surgery. #PrayersForKarlie Non-Textual and punctuation Features relating to punc- In this work, we have used the last two interaction features tuation and emoticons such as presence of ”!/?” are ex- only for comparison study, while other features are part of pected to add the discriminating qualities of a learning an extension work primarily focusing on iteration specific model. models in identifying life events. 4.2 Interaction and Social Feature 5. EXPERIMENTAL RESULT Unigram is a basic model for classification and the result In this step, we analyse the experimental steps and present shows a reasonable accuracy including a poor performance the results of classifications. We started with the ground- for the new born event. This motivated us to further ex- truth annotation process followed by classification steps and plore the feature space and extract more defining attributes their results. of an event in terms of activity and interactions based on the simple logic that important events are bound to gener- ate more attention and activity within the immediate per- 5.1 Ground Truth Annotation In the absence of any benchmark data for personal event sonal network of an individual. Accordingly, we computed detection prepared a gold standard dataset with manual an- the following Twitter specific features concerning to a tweet notation of 2 users with computing background . Annotators and the user. These features can be broadly classified into were given 1000 tweets per event for annotation. These 1000 two categories: 1) Activity and 2) Attention. Activity tweets are randomly selected from the filtered dataset. In- features (first four in the list below) are based on userÕs struction for annotation was to annotate a tweet as event activity (tweets, re-tweet and replies) while attention fea- positive (presence of event) if they consider the tweet de- tures are the measures of engagement between the user and scribes an event happening (present e.g. today) or about to his/her network (last four features in the list below) happen with certainty (e.g. 4 days to graduation) within a month’s time window. It is difficult to precisely define an event as most of the tweets are not reported exactly during 1. Tweets per day: Number of tweets per day a user posts the event but pre and post event. Since our objective is to 2. Re-tweets per day: Number of tweets per day a user identify the event from userÕs timeline with definitive time posts. stamp attached to the event, we opted for a 1 month time interval. We retained those tweets (304) as event positive 3. Replies per day: Number of replies given by the user tweets whenever both the annotators agreed on the label. to other users. It is imperative to mention that event negative tweets are simply those where annotators felt that a particular event is 4. Unique mentions per day: Number of unique mention not occurring despite the presence of event related keyword. (users addressed) in a day by the user. 5. Number of times the user is mentioned in a day 5.2 Event Detection: Unigram Model(UNI) Our first model is the simplest bag-of-word model where 6. Number of times a user is replied to, by other users word frequencies are used as features for document classifi- cation. In our case, each tweet is considered 1 document. 7. Number of times a tweet is re-tweeted by other users We first applied a String to word vector filter that coverts ** the strings into numerical features. Then we trained our model with 10-fold cross validation using four different types 8. Number of times a tweet is marked as ”favourite” by of classifiers: Naive Bayes (NB), Multinomial Naive Bayes other users.** (MNB), Support Vector Machine (SVM) and Decision Tree cial relation terms( my friend, sister etc.), temporal terms (today, week, morning etc.), sentiment strength of a tweet. POS tagging was done using Stanford tagger6 and sentiment was derived using the Sentistrength java library[?]. Recognizing Temporal Expression:Temporal features tend to be implicit, diverse, and informal (e.g. last week, hourly, around the corner). Identifying these references within the vicinity of an event term occurrence increases the likeli- hood of accurate detection. Moreover, we need to resolve the tense of the verb as well to know weather the tweet is about some future event, or past. In this paper, we are using the time terms of LIWC dictionary which has 68 time inducing Figure 1: AUC curve for different events. terms (e.g. forever,week,until etc.). This feature also used as a binary feature in the second classification model. (J48) implemented in machine learning library Weka [?]. We Average accuracy of the second model showed an average im- evaluated our model on the test set (100 from each event) provement of 4-5 % in precision score over the initial model and performance of these classifiers reported in terms of Re- for all the events, showing that simple lexical features are call (is the number of correct results divided by the number able to capture some of the diversity. For brevity purpose of results that should have been returned) Precision (is the we are only showing the results of the top classifier (SVM). number of correct results divided by the number of all re- turned results) and F-score (harmonic mean). Table 3 (fig. 2) shows the average precision, recall and F score for all the Table 4: Precision, Recall and F-measure for events. However SVM performed best in 4 out of 5 followed (UNI+META) Model (SVM). by Naive Bayes. Graduation (.8) has highest precision score whereas ”New job” has the highest recall (.95) score. The Event Precision Recall F-Measure most difficult event is the ”New born” across all the classifiers Graduation 0.83 0.81 0.819 with lowest precision score (.55). Marriage 0.77 0.83 0.798 New Job 0.818 0.93 0.865 Examining the ROC curves which plots the true positives New Born 0.61 0.92 0.733 (TP) vs false positives(FP) and indicates the area under Surgery 0.77 0.87 0.816 curve (figure 1) (AUC: probability that a classifier will rank a randomly chosen positive instance higher than a randomly chosen negative example) ranges from .71 to .75 giving a reasonable quality of the learners. NB performs better than 5.4 Event Detection: Model with Interaction SVM with an average of .77 against .72 across all events. Features (UNI+META+INT) Inherent in social media and social networks, it is intuitive to hypothesise that interesting events will stimulate inter- Table 3: Average precision, recall and f-Measure esting and increased interaction among the friend circle of from all classifiers based on unigram model. the user in the form of replies and sharing. The third and the final model takes advantage of these interaction features Event Precision Recall F-Measure embedded in microblogging sites through mechanisms like Graduation 0.80 0.80 0.73 retweet and favourites. Each tweet is now represented with Marriage 0.75 0.87 0.79 two more features besides the above lexical features for clas- New Job 0.78 0.95 0.80 sification. We used only SVM as the classifier because of its New Born 0.55 0.92 0.68 superior performance in previous two occasions. Results of Surgery 0.72 0.87 0.76 the final model (table 5) are reported by means of precision score per event. A final comparison of four models (UNI, UNI+META, UNI+META+INT and INT) is shown in fig- Analysis of error classification mainly showed the diversity of ure 3. The result shows that, although the hybrid model language constructs among the misclassified tweets. Since performed better than the unigram-based one (UNI), the the model is purely content based, any variation not cap- improvement was marginal. On the other hand, the model tured by the model are missed from the result. based only on interaction features (INT) performed worst, where accuracy dropped to 53-61%. . 5.3 Event Detection: Model with Contextual Lexical Patterns (UNI+META) 6. CONCLUSION This paper describes event detection from personal timeline Bag-of-words or unigram model is the basic approach yet of a user in Twitter. Existing detection tasks predominantly proved to have reasonable accuracy though with lots of false focused on public events and events concerning celebrities positives. This led us to refine the model with more lexical both from news articles and social media whereas personal features and features such as sentiment. We considered fea- life events are mostly overlooked. We started with 5 life tures (described in sec. 4) such as co-occurring terms (e.g. 6 prayers, hospital for surgery), POS tagging, presence of so- http://nlp.stanford.edu/software/tagger.shtml of the 18th ACM Conference on Information and Table 5: Precision, Recall and F-measure for Knowledge Management, CIKM ’09, pages 523–532, (UNI+META+INT) Model (SVM). New York, NY, USA, 2009. ACM. [4] B. D. Eugenio, N. Green, and R. Subba. Detecting Event Precision Recall F-Measure Graduation 0.85 0.83 0.839 Life Events in Feeds from Twitter. pages 274–277. Marriage 0.79 0.83 0.809 Ieee, 2013. New Job 0.82 0.91 0.862 [5] C. S. Firan, M. Georgescu, W. Nejdl, and R. Paiu. New Born 0.64 0.92 0.754 Bringing order to your photos: Event-driven classification of flickr images based on social Surgery 0.78 0.87 0.822 knowledge. In Proceedings of the 19th ACM International Conference on Information and Knowledge Management, CIKM ’10, pages 189–198, New York, NY, USA, 2010. ACM. [6] J. Gl§ck and S. Bluck. Looking back across the life span: A life story account of the reminiscence bump. Springer, 2007. [7] Q. He, K. Chang, and E.-P. Lim. Analyzing feature trajectories for event detection. In Proceedings of the 30th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’07, pages 207–214, New York, NY, USA, 2007. ACM. [8] A. Jackoway, H. Samet, and J. Sankaranarayanan. Identification of live news events using Twitter. In Figure 2: A comparative performance of four differ- book1, page 1, New York, New York, USA, 2011. ACM ent models. Press. [9] A. Java, X. Song, T. Finin, and B. Tseng. Why we events and trained 5 different binary classifiers based on twitter: Understanding microblogging usage and bag-of-word features which gave 55 to 80% precision on a communities. In Proceedings of the 9th WebKDD and test dataset with an average AUC of 77%. The learning 1st SNA-KDD 2007 Workshop on Web Mining and models were further streamlined with meta features such as Social Network Analysis, WebKDD/SNA-KDD ’07, sentiment, temporal, social relation terms, emoticons and pages 56–65, New York, NY, USA, 2007. ACM. punctuations features, which improved the classification per- [10] S. Papadopoulos, C. Zigkolis, Y. Kompatsiaris, and formance by 4-5%, however addition of interaction feature A. Vakali. Cluster-based landmark and event detection in the third classifier did not yield substantial improvement for tagged photo collections. In book1, volume 18, contrary to the expectation. This final result is a stronger pages 52–63, Los Alamitos, CA, USA, Jan. 2011. motivation for an in-depth analysis of these features in our IEEE Computer Society Press. future work. We also aimed to adopt an unsupervised ap- [11] S. Phuvipadawat and T. Murata. Breaking news proach to detect life events as there may be many more un- detection and tracking in twitter. In Proceedings of the expected events happening in one’s life bearing substantial 2010 IEEE/WIC/ACM International Conference on influence in life and eligible to be included . Web Intelligence and Intelligent Agent Technology - Volume 03, WI-IAT ’10, pages 120–123, Washington, 7. ACKNOWLEDGMENT DC, USA, 2010. IEEE Computer Society. This work was supported by EPSRC project ReelLives [12] T. Sakaki. Earthquake shakes twitter users : (EP/L004062/1). Real-time event detection by social sensors. In Proceedings of the 19th International Conference on 8. REFERENCES World Wide Web, 2009. [1] P. Agarwal, R. Vaithiyanathan, S. Sharma, and [13] M. Thelwall, K. Buckley, G. Paltoglou, and D. Cai. G. Shroff. Catching the Long-Tail : Extracting Local Sentiment strength detection in short informal text, News Events from Twitter. In book1, pages 379–382, 2010. 2012. [14] C. L. Wayne. Topic detection tracking (tdt). In In [2] E. Benson, A. Haghighi, and R. Barzilay. Event Proceedings DARPA Broadcast News Transcription Discovery in Social Media Feeds. In book1, volume 3, and Understanding Workshop, page 98, 1998. pages 389–398. Association for Computational [15] J. Weng, Y. Yao, E. Leonardi, F. Lee, and B.-s. Lee. Linguistics, 2011. Event detection in twitter. In book1, pages 401–408. [3] L. Chen and A. Roy. Event detection from flickr data Ieee, 2011. through wavelet-based spatial analysis. In Proceedings