{
"paper_id": "I13-1035",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T07:15:25.419937Z"
},
"title": "Feature-Rich Segment-Based News Event Detection on Twitter",
"authors": [
{
"first": "Yanxia",
"middle": [],
"last": "Qin",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harbin Institute of Technology",
"location": {
"postCode": "150001",
"settlement": "Harbin",
"country": "China"
}
},
"email": "yxqin@mtlab.hit.edu.cn"
},
{
"first": "Yue",
"middle": [],
"last": "Zhang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Singapore University of Technology and Design",
"location": {
"postCode": "138682",
"country": "Singapore"
}
},
"email": "yue_zhang@sutd.edu.sg"
},
{
"first": "Min",
"middle": [],
"last": "Zhang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harbin Institute of Technology",
"location": {
"postCode": "150001",
"settlement": "Harbin",
"country": "China"
}
},
"email": "zhangminmt@hotmail.com"
},
{
"first": "Dequan",
"middle": [],
"last": "Zheng",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Harbin Institute of Technology",
"location": {
"postCode": "150001",
"settlement": "Harbin",
"country": "China"
}
},
"email": "dqzheng@mtlab.hit.edu.cn"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "Event detection on Twitter is an important and challenging research topic. On the one hand, Twitter provides first-hand information and fast broadcasting. On the other, challenges include short and noisy content, big volume data and fast-changing topics. Dominant approaches for Twitter event detection model events by clustering tweets, words or segments, while segments have been proven to be advantageous over both words and tweets in news event detection. We study segment-based news event detection, for which existing heuristic-based methods suffer from low recall. We propose feature-based event filtering to address this issue. Our filter incorporate a rich family of features that are empirically proven to be valuable. Experimental results show that our event detection system outperforms the state-of-theart baseline with doubled recall and increased precision.",
"pdf_parse": {
"paper_id": "I13-1035",
"_pdf_hash": "",
"abstract": [
{
"text": "Event detection on Twitter is an important and challenging research topic. On the one hand, Twitter provides first-hand information and fast broadcasting. On the other, challenges include short and noisy content, big volume data and fast-changing topics. Dominant approaches for Twitter event detection model events by clustering tweets, words or segments, while segments have been proven to be advantageous over both words and tweets in news event detection. We study segment-based news event detection, for which existing heuristic-based methods suffer from low recall. We propose feature-based event filtering to address this issue. Our filter incorporate a rich family of features that are empirically proven to be valuable. Experimental results show that our event detection system outperforms the state-of-theart baseline with doubled recall and increased precision.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "We study news event detection from Twitter messages (tweets). Generally, tweets can be classified into three groups: 1) news events, or breaking news such as \"Manchester united Vs Athletic in Jan. 1st\"; 2) hot topics that spread among a large amount of Twitter users, such as horoscope topics (e.g. \"You have recently experienced a phase of expansionism and it's... More for Sagittarius\"); and 3) heterogeneous collections or, meaningless non-event tweets, such as \"Need buddy wanna chat\". Some previous work (Cataldi et al., 2010; Kasiviswanathan et al., 2011; Diao et al., 2012) regards both news events and hot topics as subjects of detection, while other work (Jackoway et al., 2011; Sakaki et al., 2010; Becker et al., 2012) only detects news events. We are interested in the latter, for which most previous work detects only specific types of events. For example, Sakaki et al. (2010) detect earthquake events from Twitter. In this paper, we study event detection in general.",
"cite_spans": [
{
"start": 509,
"end": 531,
"text": "(Cataldi et al., 2010;",
"ref_id": "BIBREF4"
},
{
"start": 532,
"end": 561,
"text": "Kasiviswanathan et al., 2011;",
"ref_id": "BIBREF9"
},
{
"start": 562,
"end": 580,
"text": "Diao et al., 2012)",
"ref_id": "BIBREF6"
},
{
"start": 664,
"end": 687,
"text": "(Jackoway et al., 2011;",
"ref_id": "BIBREF8"
},
{
"start": 688,
"end": 708,
"text": "Sakaki et al., 2010;",
"ref_id": "BIBREF17"
},
{
"start": 709,
"end": 729,
"text": "Becker et al., 2012)",
"ref_id": "BIBREF2"
},
{
"start": 870,
"end": 890,
"text": "Sakaki et al. (2010)",
"ref_id": "BIBREF17"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Compared with event detection in news texts, Twitter provides more opportunities and challenges. Yom-Tov and Diaz (2011) report that Twitter can broadcast news faster than traditional media, which provides an opportunity for event detection in Twitter. On the other hand, there are challenges in event detection from Twitter data: 1) tweets are too short and sometimes cannot carry enough information; 2) tweets contain many noisy words, which can be harmful for event detection and 3) the volume of Twitter data is very large, which makes event detection a big data problem.",
"cite_spans": [
{
"start": 97,
"end": 120,
"text": "Yom-Tov and Diaz (2011)",
"ref_id": "BIBREF19"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The dominant approach for Twitter event detection is clustering. Similar tweets (Becker et al., 2011; Li et al., 2012c) or words (Platakis et al., 2009; Lee et al., 2012) are group into a cluster, before clusters are classified into either news events or non-events. A recent paper (Li et al., 2012a) showed that segments (i.e. ngrams; see Section 2) can be advantageous over both tweets and words for clustering. As segments have much smaller quantity than tweets and are more semantically meaningful than words, they are better units to be clustered. We take Li's system (Twevent) as our baseline system. Twevent apply a heuristic-based method (newsworthiness) to filter out hot topics and heterogenous clusters from news events. Newsworthiness is calculated by similarity between edges in a cluster, and whether segments of the cluster frequently appear in Wikipedia. Both similarity of edges and Wikipedia are useful in filtering out heterogeneous collections from news events, while Wikipedia can also separate some news and topics. However, there are several problems with this approach: 1) newsworthiness cannot distinguish news from some topics, includes horoscope topics (\"sagittarius; approach; big trouble\") and topics such as \"hitler; fox; megan fox; rip; megan; selena gomez\", which contain segments that can also frequently occur in Wikipedia; 2) as a single measure, newsworthiness is subject to a tradeoff between precision and recall, while a high precision can be obtained only with an extremely low recall (about 10%).",
"cite_spans": [
{
"start": 80,
"end": 101,
"text": "(Becker et al., 2011;",
"ref_id": "BIBREF1"
},
{
"start": 102,
"end": 119,
"text": "Li et al., 2012c)",
"ref_id": "BIBREF13"
},
{
"start": 129,
"end": 152,
"text": "(Platakis et al., 2009;",
"ref_id": "BIBREF14"
},
{
"start": 153,
"end": 170,
"text": "Lee et al., 2012)",
"ref_id": "BIBREF10"
},
{
"start": 282,
"end": 300,
"text": "(Li et al., 2012a)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "On the other hand, tweets contain useful information that can address the weakness of newsworthiness. For example the \"Follow spree\" topic, which refers to following-back activities by celebrities to their fans, can be recognized by the common hashtag suffix \"followspree\". Another example is that news tweets are more likely to contain url links. We propose a classifier based method for event filtering and define a set of novel features that capture statistical, social and textual information from event clusters. Some of the features are useful in getting rid of heterogeneous collections while others may be useful for recognizing news events from hot topics. We call our system Feature-Rich segment-based news Event Detection system on Twitter (FRED). Experimental results show that our system FRED outperforms the state-of-the-art event detection system Twevent by significantly increased precision and doubled recall.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In this section, we introduce the segment-based event detection method of Li et al. (2012a) , which consists of three steps: tweet segmentation, bursty segment detection and segment clustering. Tweet Segmentation splits tweet into non-overlapping segments, which maybe unigrams or N-grams (2-5 grams). For a certain time window, segments that show a bursty frequency pattern are selected as bursty segments. Segment clustering groups bursty segments about same event into one cluster regarding them as one event.",
"cite_spans": [
{
"start": 74,
"end": 91,
"text": "Li et al. (2012a)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Segment-Based Event Detection",
"sec_num": "2"
},
{
"text": "Tweet segmentation can be regarded as a optimization problem to partition tweet with the use of Microsoft Web N-Gram service 1 and Wikipedia 2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tweet Segmentation",
"sec_num": "2.1"
},
{
"text": "1 http://web-ngram.research.microsoft.com/info/ 2 http://www.wikipedia.org/ The objective function is defined as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tweet Segmentation",
"sec_num": "2.1"
},
{
"text": "arg max s 1 ,...,sm C(d) = m \u2211 i=1 C(s i ) (1)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tweet Segmentation",
"sec_num": "2.1"
},
{
"text": "where d is a tweet from Tweet stream, {s 1 , ..., s m } are the segments in tweet d and C is the function which measures the stickiness of a tweet or segment. In particular:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tweet Segmentation",
"sec_num": "2.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "C(s) = L(s) \u2022 e Q(s) \u2022 S(scp(s))",
"eq_num": "(2)"
}
],
"section": "Tweet Segmentation",
"sec_num": "2.1"
},
{
"text": "where the length L(s) is defined in Eq 3, and a longer s makes it typically less sticky. Q(s) is the probability that s appears as an anchor text in Wikipedia articles; frequently-appearing anchor texts are more semantically meaningful. S(\u2022) is the sigmoid funcion. scp(s) is a cohesiveness measurement of segment s defined with symmetric conditional probability, as shown in Eq 4. Better combination of words when forming segments leads to higher cohesiveness value.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tweet Segmentation",
"sec_num": "2.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "L(s) = { |s|\u22121 |s| , f or |s| > 1 1, f or |s| = 1 (3) scp(s) = log P r(s) 2 1 n\u22121 n\u22121 \u2211 i=1 P r(w i 1 )P r(w n i+1 )",
"eq_num": "(4)"
}
],
"section": "Tweet Segmentation",
"sec_num": "2.1"
},
{
"text": "In the above equations, a segment s can be written as {w 1 . . . w n }(n > 1), where P r(\u2022) is the prior probability derived from the Microsoft Web N-gram service.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Tweet Segmentation",
"sec_num": "2.1"
},
{
"text": "From the large number of segments resulting from the last step, a small portion of bursty ones are selected for event clustering since segments with a burst frequency are more representative for a breaking news in the data stream. For convenience, we take a time window t as the time unit for bursty segment detection and segment clustering. N t refers to the number of tweets within the time window t, and f s,t represents the number of tweets that contain segment s within t. If f s,t > E [s|t] , then a segment s is a bursty segment. E[s|t] is the expected number of tweets that contain s within t. As N t is sufficiently large, the Gaussian distribution is used to model the probability of f s,t .",
"cite_spans": [
{
"start": 491,
"end": 496,
"text": "[s|t]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Bursty Segment Detection",
"sec_num": "2.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "P (f s,t ) \u223c N (N t p s , N t p s (1 \u2212 p s ))",
"eq_num": "(5)"
}
],
"section": "Bursty Segment Detection",
"sec_num": "2.2"
},
{
"text": "where p s is expected probability of tweets contain s, calculated as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Bursty Segment Detection",
"sec_num": "2.2"
},
{
"text": "p s = 1 L L \u2211 t=1 f s,t N t (6)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Bursty Segment Detection",
"sec_num": "2.2"
},
{
"text": "L is number of time windows containing s. E [s|t] = N t p s . Even after filtering out nonbursty segments, a large amount of bursty segments remain. Bursty weight w b (s, t) is assigned to each bursty segments and the top K bursty segments are chosen for further processing. K is set to",
"cite_spans": [
{
"start": 44,
"end": 49,
"text": "[s|t]",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Bursty Segment Detection",
"sec_num": "2.2"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u221a N t in Twevent. w b (s, t) = P b (s, t)log(u s,t )",
"eq_num": "(7)"
}
],
"section": "Bursty Segment Detection",
"sec_num": "2.2"
},
{
"text": "P b (s, t) is the bursty probability and u s,t means the user frequency of s helping to filter out some noisy segments, as the more users talk about the segment s, the more popular and meaningful it is. u s,t is calculated as the number of users who post tweets containing s within t.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Bursty Segment Detection",
"sec_num": "2.2"
},
{
"text": "P b (s, t) = S(10 \u00d7 f s,t \u2212 (E[s|t] + \u03c3[s|t]) \u03c3[s|t] ) (8) \u03c3[s|t] = \u221a N t p s (1 \u2212 p s )",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Bursty Segment Detection",
"sec_num": "2.2"
},
{
"text": "is the standard deviation of Gaussian distribution in Eq 5.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Bursty Segment Detection",
"sec_num": "2.2"
},
{
"text": "k-Nearest Neighbor graph (kNNgraph) clustering method, is applied to group bursty segments into clusters. The kNNgraph clustering method takes a complete graph of bursty segments with edges representing similarity between segments as input and output event clusters. It groups two segments into same cluster only when they are in each other's k-nearest neighbors. k is a key parameter to control the size of clusters. We choose value for k in Section 4. The output of kNNgraph clustering is an event cluster set corresponding to the time window t, denoted as Gset(t). All Gset(t) sets are gathered to a whole event cluster set Gset. Gset is manually labeled for further use, introduced in Section 4.2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Segment Clustering",
"sec_num": "2.3"
},
{
"text": "Temporal features and text similarity are incorporated when calculating similarity between two segments s 1 , s 2 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Segment Clustering",
"sec_num": "2.3"
},
{
"text": "sim t (s 1 , s 2 ) = M \u2211 m=1 w t (s 1 , m)w t (s 2 , m)sim(T 1 , T 2 ) (9) < t 1 . . . t M >",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Segment Clustering",
"sec_num": "2.3"
},
{
"text": "are M sub time windows of the time window t. Frequency of segment s in the sub time window t m is denoted as f t (s, m). w t (s, m) is the frequency weight of s in t m , which serves as a temporal feature and is shown in Eq 10. T i denotes a set of tweets containing s i within t m . sim(T 1 , T 2 ) measures text similarity between the two sets of tweets T 1 , T 2 . Tweets in T i are concatenated as a pseudo document, and cosine similarity is applied for calculating distance. Pseudo documents are represented by the Vector Space Model, weighted by TF-IDF. TF value is the number of tweets containing word w within the sub time window t m and DF value is the number of tweets containing w in the whole twitter corpus.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Segment Clustering",
"sec_num": "2.3"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "w t (s, m) = f t (s, m) M \u2211 m \u2032 =1 f t (s, m \u2032 )",
"eq_num": "(10)"
}
],
"section": "Segment Clustering",
"sec_num": "2.3"
},
{
"text": "3 Feature-Rich Event Filter",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Segment Clustering",
"sec_num": "2.3"
},
{
"text": "The clusters in the kNNgraph clustering result Gset contain news events, hot topics and heterogeneous clusters, corresponding to the three types of tweets mentioned in the Introduction. Our goal, which is to recognize news events from other two types of event clusters, is a challenging task because hot topics and news events can both have bursty frequency and share similar characteristics.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Segment Clustering",
"sec_num": "2.3"
},
{
"text": "Twevent utilizes a heuristic-based method for event filtering using information from Wikipedia. A heuristic equation, newsworthiness, is used to determine whether an event cluster is a news event or not, whereas all clusters with a high newsworthiness score is news events. The newsworthiness \u00b5(e) of an event cluster e containing segment set S e and edge set G e is calculated as follows.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Event Filter in Twevent",
"sec_num": "3.1"
},
{
"text": "EQUATION",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [
{
"start": 0,
"end": 8,
"text": "EQUATION",
"ref_id": "EQREF",
"raw_str": "\u00b5(e) = \u2211 s\u2208Se \u00b5(s) |S e | \u2022 \u2211 g\u2208Ge sim(g) |S e |",
"eq_num": "(11)"
}
],
"section": "Event Filter in Twevent",
"sec_num": "3.1"
},
{
"text": "where \u00b5(s) of segment s is calculated as:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Event Filter in Twevent",
"sec_num": "3.1"
},
{
"text": "\u00b5(s) = max l\u2208s e Q(l) \u2212 1 (12)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Event Filter in Twevent",
"sec_num": "3.1"
},
{
"text": "l is sub-phrase of s and Q(l) is the probability that l appears as anchor text in Wikipedia articles. An event cluster e is taken as a news event only if it satisfies the condition that \u00b5 max /\u00b5(e) < \u03c4 , where \u03c4 is a threshold for newsworthiness, \u00b5 max is the maximum of \u00b5(e) in time window t. Lower \u03c4 value leads to high precision and low recall, which is the limitation of newsworthiness. Rich information from tweets and clusters themselves can be useful in alleviating this problem.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Event Filter in Twevent",
"sec_num": "3.1"
},
{
"text": "In order to incorporate rich features, we take event filtering as a binary classification problem, where class 'T' means true news event class includes news events and class 'F' represents false news event class containing hot topics and heterogenous clusters. In our filter, event clusters in Gset are represented with a set of cluster-level features, and classified into T or F by a SVM 3 classifier. All clusters in class T, represented as Eset, form the final news event result. Features used to represent event clusters are shown in Section 3.3.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Event Filter in FRED",
"sec_num": "3.2"
},
{
"text": "We collect three types of features for the filter, representing statistical, social and textual information related of event clusters, respectively. Some of the features are designed to filter out heterogeneous clusters, while others to distinguish news events from hot topics. Given an event cluster e and the corresponding time window t (from which e is extracted), we have the following information: 1) Gset(t), a sub set of Gset corresponding to t. 2) S e , the set of segments in e and G e , the set of edges in e. 3) T (e), which consists of tweets that are related to e containing at least one segment of S e and being posted in t. 4) relU (e), which represents users who posted the tweets in relT (e), and U t , which denotes the number of users who published tweets within t.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Features",
"sec_num": "3.3"
},
{
"text": "For statistical features, we collect direct statistical information from event clusters, such as how many segments and edges it contains, the density of the event graph and so on.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "\u2022 seg, which refers to the segment number of e, calculated as |S e |/ max e \u2032 \u2208Gset(t)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "(S e \u2032 ). News events and hot topics contain more segments than heterogeneous clusters generally.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "\u2022 edge, which refers to number of the edges of e, defined as G e / max e \u2032 \u2208Gset(t)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "(G e \u2032 ). Similar with segment number, heterogeneous clusters usually have less edges than news and topics.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "\u2022 wiki, the average of newsworthiness for all segments in S e . A higher wiki value indicates that the event cluster contains more meaningful and important segments. wiki is able to distinguish news events from hot topics and heterogeneous clusters, as shown Li et al. (2012a) .",
"cite_spans": [
{
"start": 259,
"end": 276,
"text": "Li et al. (2012a)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "\u2022 dup is designed to filter out some specific heterogeneous clusters that contains words sharing the same lemma. For example the event cluster e7 in Table 3 , which words sharing lemma \"feel\". dup can be obtained by stemming all unigrams appeared in S e and calculating the number of duplicated stemmed unigrams out of all stemmed unigrams.",
"cite_spans": [],
"ref_spans": [
{
"start": 149,
"end": 156,
"text": "Table 3",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "\u2022 sim, which refers to average similarity of all edges in G e . A bigger sim means that the event cluster is more dense, or sticky.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "\u2022 df, which refers to the number of tweets related to e out of all tweets published in t, namely |relT (e)|/N t. df could help to eliminate heterogeneous clusters, which are published by less users in less tweets.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "\u2022 udf, which refers to |relU (e)|/U t. The influence of udf is similar with df.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Statistical Features",
"sec_num": null
},
{
"text": "Tweets in relT (e) contain rich Twitter-specific social information, which may reveal the difference between news events and hot topics. For example, the more mentions (@username) exist in relT (e), the more likely e is a topic.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Social Features",
"sec_num": null
},
{
"text": "\u2022 rt, which represent how many tweets in relT (e) are retweeted. Retweet is a forwarding action on a tweet published by other users indicating an interest to the tweet. A retweeted tweet is denoted by a prefix of \"RT @username\". Retweet functions as a means of sharing and spreading without commenting to show user's opinion. A news event may have a larger fraction of retweeted tweets than others as users want to spread the news.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Social Features",
"sec_num": null
},
{
"text": "\u2022 men, which refers to the normalized number of tweets containing mention (e.g., @username) in relT (e) specifying one target receiver of tweets (e.g., \"@justinbieber\").",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Social Features",
"sec_num": null
},
{
"text": "Mention actions occur more frequently in hot topics than in news events, as users prefer showing their opinion about this topic rather than just spreading it.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Social Features",
"sec_num": null
},
{
"text": "\u2022 rep, which refers to the normalized number of reply tweets in relT (e). Reply means commenting, and a reply tweet is started with a mention. Similar with mention, reply has strong indication of conversation, and are more related to topics than news.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Social Features",
"sec_num": null
},
{
"text": "\u2022 url, which refers to the normalized number of tweets containing url link in relT (e). Url shows extra information for tweet. News events contains more information than a topic, which may not be fully expressed in a short tweet, and hense url links are likely used to refer to the original article.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Social Features",
"sec_num": null
},
{
"text": "\u2022 tag, which refers to the normalized number of tweets containing hashtag in relT (e). A hashtag (#gamecocks) is a short description of what's happening. Generally a popular hashtag indicates a hot topic or an event (e.g., \"The game got a little exciting today but we got the win! #gamecocks\").",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Social Features",
"sec_num": null
},
{
"text": "\u2022 pst, which measures how many tweets contain words in past tense in relT (e), normalized by |relT (e)|. News events are more likely to be described formally and with more words in past tense.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Social Features",
"sec_num": null
},
{
"text": "Besides above groups of features, text information embedded in hashtag content are another valuable source of information. News events will more likely have common hashtags. For example, many tweets about \"National Football League\" games have a common hashtag \"#NFL\". Twitter topic can have common prefixes or suffixes of hashtag. For example the \"Follow spree\" topic, which is mentioned earlier, may have a common hashtag suffix \"followspree\".",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Textual Features",
"sec_num": null
},
{
"text": "\u2022 fTag, which represents how many hashtags appear in relT (e) are frequent hashtags. We extract a frequent hashtag list from whole Twitter data set by takeing the top 2000 most frequently used hashtags.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Textual Features",
"sec_num": null
},
{
"text": "\u2022 psfx. To obtain frequent hashtag prefixes/suffixes, we first filter out prefixes/suffixes of all hashtags in the data that satisfy at least one of the following conditions: 1) less than 3 characters, 2) composed by repeating one character, 3) frequency lower than 200. After arranging the prefixes in alphabetical order, we keep only the longest prefixes for the same prefix pattern. Prefixes are ranked by frequency, and the top 2000 are taken as frequent hashtag prefixes. Similarly, we could extract 2000 most frequent hashtag suffixes. pfx and sfx are used to indicate how many hashtags tweets in relT (e) contain frequent prefixes or suffixes respectively. psfx is the combination of pfx and sfx by multiplying them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Textual Features",
"sec_num": null
},
{
"text": "The Twitter data we use were crawled from Twitter timeline, which is the real-time tweet stream containing all tweets published by Twitter users from January 1st to January 15th, 2013. After removing stops words, filtering out non-English tweets and null content tweets, the data set contains 31,097,528 tweets published by 16,331,133 users with 382,475 words. Wikipedia data is used as an extra resource in the tweet segmentation tasks (Section 2.1) and event filtering (Section 3). We use the Wikipedia dump data 4 of February 4th, 2013. It includes 13,167,739 pages and 10,507,127 anchor entities that have 5 words length limit. These anchor entities' anchor probability, i.e. the number of pages that entity e appears as anchor text divided by the number of pages containing entity e, are calculated at the very beginning.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data",
"sec_num": "4.1"
},
{
"text": "We reproduced Twevent as our baseline system. Parameter \u03c4 in Twevent and gamma in FRED are tuned for best performance on a development set, which consists all event clusters on Jan. 2nd and 5th. Time window t is set to be a day and M (in Eq. 9) is 12. k in kNNgraph clustering method is set to be 5, as a tradeoff of the number of event clusters and average number of segments in clusters. \u03c4 in Twevent is tuned and set to be 2 and gamma in LibSVM of FRED is 5. 10-fold cross validation is We built a standard gold set for FRED after labeling the event cluster set Gset, which is the output of the segment-based event detection (Section 2). The labeling method is shown as follows. Given an event cluster e, the segments in e and the corresponding time window t, we use the segments and t to determine whether e is related to a news. Google and Twitter search are used to assist mannual annotations of events. As a result, 4249 event clusters in Gset were manually labeled into 804 news events and 3445 nonevents. Note that some news events in Gset may be sub events of one event. For example \"The Golden Globe Awards ceremony 2013\" happened in January 13th are detected more than once, as people talked about winners for different awards. We have not merged these sub events in this paper, which will be considered for future work.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Settings",
"sec_num": "4.2"
},
{
"text": "With the event cluster set Gset, we use the precision, recall and F1-measure to evaluate the performances of FRED and Twevent, where precision is defined the fraction of news events in system-generated 'T' class event clusters (Eset for FRED), and recall measures how many manually labeled news events are detected out of all news events in Gset. F1 measure is calculated for an overall evaluation. Note that given our annotations, which is much larger than that of Li (Li et al., 2012a) , we can give a better estimation of recall, which Li et al. were not able to report in detail (they used the number of detected news as recall, which did not reveal the real recall notion).",
"cite_spans": [
{
"start": 469,
"end": 487,
"text": "(Li et al., 2012a)",
"ref_id": "BIBREF11"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Settings",
"sec_num": "4.2"
},
{
"text": "As we show some statistical results of tweet segmentation (Section 2.1), we obtained 1,604,129 distinct segments with 22.3% unigrams, 72% 2grams and 5.7% 3-5 grams.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Experimental Results and Analysis",
"sec_num": "4.3"
},
{
"text": "In Table 1 we show the results of feature ablation test. ExpID is the experiment id. FeatureSet is the features we used for current experiment. All means all statistical, social and textual features. Diff means the difference between F1 in current experiment and experiment 0, and a smaller Diff indicates that the feature is more valuable.",
"cite_spans": [],
"ref_spans": [
{
"start": 3,
"end": 10,
"text": "Table 1",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Effectiveness of Features",
"sec_num": null
},
{
"text": "The experimental results show that nearly all features contribute to event filter on either precision or recall. Features can be partitioned into three groups according to their impact on precision and recall: 1) features that are useful only for precision include df, udf, dup, rt. 2) features that are useful for both precision and recall includes rest of features such as wiki, sim, url etc. 3) feature pst is slightly harmful for precision and recall.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Effectiveness of Features",
"sec_num": null
},
{
"text": "The most valuable features to our system are wiki, sim, rep, url and tag. wiki is extra resource obtained from Wikipedia, and contributes to valuable segments in event clusters. sim indicates denser event cluster with stronger connections between segments, while replied tweet number, url number and hashtag number are social features embedded in tweets related to event clusters. Results show there are bigger differences in these features between news events and others clusters when compared to other social features. The experimental results of FRED and baseline systems are presented in Table 2 . Twevent u is a variant of Twevent, which uses unigrams (words) instead of segments in the event detection. #Evt is the number of news events.",
"cite_spans": [],
"ref_spans": [
{
"start": 592,
"end": 599,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Effectiveness of Features",
"sec_num": null
},
{
"text": "The experimental result of Twevent (precision 75.7%) is lower than that reported by Li et al. (2012a) (precision 86.1%) . It is likely to be caused by 1) different Twitter data, Li use Singapore Twitter data containing 4.3 million tweets in one month while ours is global Twitter data of 31.1 million tweets in half a month; 2) horoscope topics are very popular in our data, which cannot be filtered out by Twevent. Because horoscope topics greatly influence the performance of Twevent, we performed a manual filtering to them for a better result. Twevent without the extra process yields 125 event clusters with a low precision of 64.8%. No extra filtering process was necessary for FRED.",
"cite_spans": [
{
"start": 84,
"end": 119,
"text": "Li et al. (2012a) (precision 86.1%)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Effectiveness of Features",
"sec_num": null
},
{
"text": "The results in Table 2 show that, 1) segments are better than words for news event detection as Twevent outperforms Twevent u , which brings in more heterogeneous collections; 2) our system FRED performs better than Twevent with significantly increased precision and doubled recall, which proves that feature-rich event filter could alleviate the low recall problem in Twevent.",
"cite_spans": [],
"ref_spans": [
{
"start": 15,
"end": 22,
"text": "Table 2",
"ref_id": "TABREF3"
}
],
"eq_spans": [],
"section": "Effectiveness of Features",
"sec_num": null
},
{
"text": "We show some example event clusters in Table 3. Lm refers to manually annotated class label. Lt and Lf refer to class labels generated by Twevent and FRED, respectively. The labeling results in Table 3 show that Twevent and FRED made different types of mistakes. As mentioned earlier, Twevent (without manual filtering) always fails to distinguish horoscope topics, while FRED can. From e2-e3 and e4-e5, we can also see that Twevent's labeling result changes for same news events while FRED gives consistent labels. Note that one important difference between FRED and Twevent is that the former uses some supervision. Preliminary experiments show that unsupervised clustering such as k-means clustering cannot effectively bring the benefits of rich-features.",
"cite_spans": [],
"ref_spans": [
{
"start": 194,
"end": 201,
"text": "Table 3",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Analysis",
"sec_num": null
},
{
"text": "Football and basketball games, which appear almost everyday, take a large fraction of news events. Events such as the 27th Golden Globes Award ceremony hosted on January 13th, show bursty frequency patterns from late January 13th to 14th. Topics such as horoscope topics are popular everyday. At least from our data, the most popular hobbies of the globe seem to be football games.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Analysis",
"sec_num": null
},
{
"text": "Among all events, concert news or gossips about celebrities such as \"Justin Bieber\" and \"Taylor Swift\" draw much more and much longer attention. For example, e4 in Table 3 , which is a news event related to Justin Bieber, continues to appear as news in many days very longer than e1 (a news related to song). New episode of TV programs and TV series such as \"Big Brother\" and \"Pretty Little Liars\" are also popular news events.",
"cite_spans": [],
"ref_spans": [
{
"start": 164,
"end": 171,
"text": "Table 3",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "Analysis",
"sec_num": null
},
{
"text": "Document-pivot clustering methods are frequently used in event detection on social media, in which short messages are regarded as documents (Becker et al., 2011; Li et al., 2012c) . Becker et al. (2011) represent text content of a tweet as a TF-IDF weight vector and apply an incremental clustering algorithm to group similar tweets with one cluster regarded as an event. In the following event classification phase, temporal, social, topical and Twitter-centric features are used to represent each cluster and clusters are determined whether they are event-related or topic-related or non-event.",
"cite_spans": [
{
"start": 140,
"end": 161,
"text": "(Becker et al., 2011;",
"ref_id": "BIBREF1"
},
{
"start": 162,
"end": 179,
"text": "Li et al., 2012c)",
"ref_id": "BIBREF13"
},
{
"start": 182,
"end": 202,
"text": "Becker et al. (2011)",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "As social media data is on an extremely big scale, document-pivot clustering methods are ineffective as they are time-and memory-consuming. In contrast, in feature-pivot clustering methods, only features (words) that show a burst frequency pattern in a time window are extracted and then clustered into groups to get events. In addition to improving clustering efficiency, detecting bursty features also plays an important role for feature selection as social media messages are very noisy.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "In most feature-pivot clustering methods, events are represented as a few representative words showing what happened, which may cause events to be difficult to understand (Li et al., 2012a; Platakis et al., 2009; Lee et al., 2012; Fung et al., 2005) . Li et al. (2012a) adopted tweet segmentation in their event detection system Tweven- Li et al. (2012b) for an named entity recognition system on Twitter. They claim that segments are much more meaningful and easier to read than words. Twevent is the most related work to this paper. We adopt tweet segmentation, and segment tweets into non-overlapping segments that are regarded as bursty feature candidates, and utilize a feature-pivot clustering method to group bursty segments into clusters as events. The difference between this paper and Twevent is that they use a simple measurement (newsworthiness) to filter out meaningless twitter topics from events, while we propose a classifier based filter to distinguish news events and twitter topics. The advantage of our system is that it supports the definition of rich features, some of which are helpful to eliminate heterogeneous clusters and others can distinguish news events and hot topics. We will explore their functions in this paper.",
"cite_spans": [
{
"start": 171,
"end": 189,
"text": "(Li et al., 2012a;",
"ref_id": "BIBREF11"
},
{
"start": 190,
"end": 212,
"text": "Platakis et al., 2009;",
"ref_id": "BIBREF14"
},
{
"start": 213,
"end": 230,
"text": "Lee et al., 2012;",
"ref_id": "BIBREF10"
},
{
"start": 231,
"end": 249,
"text": "Fung et al., 2005)",
"ref_id": "BIBREF7"
},
{
"start": 252,
"end": 269,
"text": "Li et al. (2012a)",
"ref_id": "BIBREF11"
},
{
"start": 337,
"end": 354,
"text": "Li et al. (2012b)",
"ref_id": "BIBREF12"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "In addition to the above group of work, which represents events with a few messages or features showing the topic information, some researchers try to extract structured information for events. Given a set of seed events, Benson et al. (2011) use a factor graph to extract artist and venue information of a concert event. Popescu et al. (2011) extract main entities, actions and audience opinions.",
"cite_spans": [
{
"start": 222,
"end": 242,
"text": "Benson et al. (2011)",
"ref_id": "BIBREF3"
},
{
"start": 322,
"end": 343,
"text": "Popescu et al. (2011)",
"ref_id": "BIBREF16"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "Data from social medias like Twitter are very sparse in presenting thousands of events, while some researchers mainly focus on specific types of events. Sakaki et al. (2010) detected disaster events like earthquakes and typhoons from Twitter. Pohl et al. (2012) tried to detect subevent to assist disaster management with Flickr and YouTube data. Agarwal et al. (2012) analyzed tweets containing specific keywords and report Fire-in Factory and Labor-Strike events. They have fixed query words and search for related messages from social media websites for data. The query words are challenges to define as they are vital to the quality of dataset, which will greatly influence the results. Becker et al. (2012) tried to generate queries for a planned event to relax the limitation. Our work mainly focus on news event detection problem on Twitter.",
"cite_spans": [
{
"start": 153,
"end": 173,
"text": "Sakaki et al. (2010)",
"ref_id": "BIBREF17"
},
{
"start": 243,
"end": 261,
"text": "Pohl et al. (2012)",
"ref_id": "BIBREF15"
},
{
"start": 347,
"end": 368,
"text": "Agarwal et al. (2012)",
"ref_id": "BIBREF0"
},
{
"start": 691,
"end": 711,
"text": "Becker et al. (2012)",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "Rich features have been used in other tasks in NLP, such as POS-tagging (Toutanova et al., 2003) , parsing (Zhang and Nivre, 2011) and machine translation (Chiang et al., 2009) . Our work is in line with these.",
"cite_spans": [
{
"start": 72,
"end": 96,
"text": "(Toutanova et al., 2003)",
"ref_id": "BIBREF18"
},
{
"start": 107,
"end": 130,
"text": "(Zhang and Nivre, 2011)",
"ref_id": "BIBREF20"
},
{
"start": 155,
"end": 176,
"text": "(Chiang et al., 2009)",
"ref_id": "BIBREF5"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related Work",
"sec_num": "5"
},
{
"text": "We proposed a feature-rich classifier to recognize news events for segment based event detection, defining novel statistical, social and textual features for the filter. Experiments showed the effectiveness of the method, and in particular some features such as the number of urls and hashtags. The feature-rich event filter led to significantly higher precision and doubled recall when compared to the state-of-the-art baseline system. In our experiments we observed that a news event can be detected more than once in one time window, which each appearance representing one aspects of the event. Building these sub-events into a hierarchy will be explored in the future.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Conclusion",
"sec_num": "6"
},
{
"text": "We use LibSVM http://www.csie.ntu.edu.tw/ cjlin/libsvm/ for the experiment.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "http://burnbit.com/download/235406/enwiki 20130204 pages articles xml bz2",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [
{
"text": "Yanxia Qin is supported by SRG-SUTD2012038 from Singapore University of Technology and Design and the National Natural Science Foundation of China (No. 61073130) from Harbin Institute of Technology. Yue Zhang is fully supported by SRG-SUTD2012038.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Acknowledgments",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Catching the long-tail: Extracting local news events from twitter",
"authors": [
{
"first": "Puneet",
"middle": [],
"last": "Agarwal",
"suffix": ""
},
{
"first": "Rajgopal",
"middle": [],
"last": "Vaithiyanathan",
"suffix": ""
},
{
"first": "Saurabh",
"middle": [],
"last": "Sharma",
"suffix": ""
},
{
"first": "Gautam",
"middle": [],
"last": "Shroff",
"suffix": ""
}
],
"year": 2012,
"venue": "ICWSM",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Puneet Agarwal, Rajgopal Vaithiyanathan, Saurabh Sharma, and Gautam Shroff. 2012. Catching the long-tail: Extracting local news events from twitter. In ICWSM.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Beyond trending topics: Real-world event identification on twitter",
"authors": [
{
"first": "H",
"middle": [],
"last": "Becker",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Naaman",
"suffix": ""
},
{
"first": "L",
"middle": [],
"last": "Gravano",
"suffix": ""
}
],
"year": 2011,
"venue": "Fifth International AAAI Conference on Weblogs and Social Media",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Becker, M. Naaman, and L. Gravano. 2011. Be- yond trending topics: Real-world event identifica- tion on twitter. In Fifth International AAAI Confer- ence on Weblogs and Social Media.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Identifying content for planned events across social media sites",
"authors": [
{
"first": "Hila",
"middle": [],
"last": "Becker",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Iter",
"suffix": ""
},
{
"first": "Mor",
"middle": [],
"last": "Naaman",
"suffix": ""
},
{
"first": "Luis",
"middle": [],
"last": "Gravano",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of WSDM",
"volume": "",
"issue": "",
"pages": "533--542",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hila Becker, Dan Iter, Mor Naaman, and Luis Gravano. 2012. Identifying content for planned events across social media sites. In Proceedings of WSDM, pages 533-542, Seattle, Washington, USA.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Event discovery in social media feeds",
"authors": [
{
"first": "Edward",
"middle": [],
"last": "Benson",
"suffix": ""
},
{
"first": "Aria",
"middle": [],
"last": "Haghighi",
"suffix": ""
},
{
"first": "Regina",
"middle": [],
"last": "Barzilay",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of ACL-HLT",
"volume": "",
"issue": "",
"pages": "389--398",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Edward Benson, Aria Haghighi, and Regina Barzilay. 2011. Event discovery in social media feeds. In Proceedings of ACL-HLT, pages 389-398, Portland, Oregon.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Emerging topic detection on twitter based on temporal and social terms evaluation",
"authors": [
{
"first": "Mario",
"middle": [],
"last": "Cataldi",
"suffix": ""
},
{
"first": "Luigi",
"middle": [
"Di"
],
"last": "Caro",
"suffix": ""
},
{
"first": "Claudio",
"middle": [],
"last": "Schifanella",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of KDD",
"volume": "4",
"issue": "",
"pages": "1--4",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Mario Cataldi, Luigi Di Caro, and Claudio Schifanella. 2010. Emerging topic detection on twitter based on temporal and social terms evaluation. In Proceed- ings of KDD, pages 4:1-4:10, Washington, D.C.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "001 new features for statistical machine translation",
"authors": [
{
"first": "David",
"middle": [],
"last": "Chiang",
"suffix": ""
},
{
"first": "Kevin",
"middle": [],
"last": "Knight",
"suffix": ""
},
{
"first": "Wei",
"middle": [],
"last": "Wang",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of NAACL",
"volume": "11",
"issue": "",
"pages": "218--226",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Chiang, Kevin Knight, and Wei Wang. 2009. 11,001 new features for statistical machine trans- lation. In Proceedings of NAACL, pages 218-226, Boulder, Colorado.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "Finding bursty topics from microblogs",
"authors": [
{
"first": "Qiming",
"middle": [],
"last": "Diao",
"suffix": ""
},
{
"first": "Jing",
"middle": [],
"last": "Jiang",
"suffix": ""
},
{
"first": "Feida",
"middle": [],
"last": "Zhu",
"suffix": ""
},
{
"first": "Ee-Peng",
"middle": [],
"last": "Lim",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of ACL",
"volume": "",
"issue": "",
"pages": "536--544",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Qiming Diao, Jing Jiang, Feida Zhu, and Ee-Peng Lim. 2012. Finding bursty topics from microblogs. In Proceedings of ACL, pages 536-544, Jeju Island, Korea.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "Parameter free bursty events detection in text streams",
"authors": [
{
"first": "Gabriel Pui Cheong",
"middle": [],
"last": "Fung",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [
"Xu"
],
"last": "Yu",
"suffix": ""
},
{
"first": "Philip",
"middle": [
"S"
],
"last": "Yu",
"suffix": ""
},
{
"first": "Hongjun",
"middle": [],
"last": "Lu",
"suffix": ""
}
],
"year": 2005,
"venue": "Proceedings of VLDB",
"volume": "",
"issue": "",
"pages": "181--192",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Gabriel Pui Cheong Fung, Jeffrey Xu Yu, Philip S. Yu, and Hongjun Lu. 2005. Parameter free bursty events detection in text streams. In Proceedings of VLDB, pages 181-192, Trondheim, Norway.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Identification of live news events using twitter",
"authors": [
{
"first": "Alan",
"middle": [],
"last": "Jackoway",
"suffix": ""
},
{
"first": "Hanan",
"middle": [],
"last": "Samet",
"suffix": ""
},
{
"first": "Jagan",
"middle": [],
"last": "Sankaranarayanan",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of LBSN",
"volume": "",
"issue": "",
"pages": "25--32",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Alan Jackoway, Hanan Samet, and Jagan Sankara- narayanan. 2011. Identification of live news events using twitter. In Proceedings of LBSN, pages 25-32, Chicago, Illinois.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "Emerging topic detection using dictionary learning",
"authors": [
{
"first": "Prem",
"middle": [],
"last": "Shiva Prasad Kasiviswanathan",
"suffix": ""
},
{
"first": "Arindam",
"middle": [],
"last": "Melville",
"suffix": ""
},
{
"first": "Vikas",
"middle": [],
"last": "Banerjee",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Sindhwani",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of CIKM",
"volume": "",
"issue": "",
"pages": "745--754",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shiva Prasad Kasiviswanathan, Prem Melville, Arindam Banerjee, and Vikas Sindhwani. 2011. Emerging topic detection using dictionary learning. In Proceedings of CIKM, pages 745-754, Glasgow, Scotland, UK.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Bursty event detection from text streams for disaster management",
"authors": [
{
"first": "Sungjun",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Sangjin",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Kwanho",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Jonghun",
"middle": [],
"last": "Park",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of WWW Companion",
"volume": "",
"issue": "",
"pages": "679--682",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Sungjun Lee, Sangjin Lee, Kwanho Kim, and Jonghun Park. 2012. Bursty event detection from tex- t streams for disaster management. In Proceedings of WWW Companion, pages 679-682, Lyon, France.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "Twevent: segment-based event detection from tweets",
"authors": [
{
"first": "Chenliang",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Aixin",
"middle": [],
"last": "Sun",
"suffix": ""
},
{
"first": "Anwitaman",
"middle": [],
"last": "Datta",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of CIKM",
"volume": "",
"issue": "",
"pages": "155--164",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chenliang Li, Aixin Sun, and Anwitaman Datta. 2012a. Twevent: segment-based event detection from tweets. In Proceedings of CIKM, pages 155- 164, Maui, Hawaii, USA.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Twiner: named entity recognition in targeted twitter stream",
"authors": [
{
"first": "Chenliang",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Jianshu",
"middle": [],
"last": "Weng",
"suffix": ""
},
{
"first": "Qi",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Yuxia",
"middle": [],
"last": "Yao",
"suffix": ""
},
{
"first": "Anwitaman",
"middle": [],
"last": "Datta",
"suffix": ""
},
{
"first": "Aixin",
"middle": [],
"last": "Sun",
"suffix": ""
},
{
"first": "Bu-Sung",
"middle": [],
"last": "Lee",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of SIGIR",
"volume": "",
"issue": "",
"pages": "721--730",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Chenliang Li, Jianshu Weng, Qi He, Yuxia Yao, Anwitaman Datta, Aixin Sun, and Bu-Sung Lee. 2012b. Twiner: named entity recognition in target- ed twitter stream. In Proceedings of SIGIR, pages 721-730, Portland, Oregon, USA.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Tedas: A twitter-based event detection and analysis system",
"authors": [
{
"first": "Rui",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Kin",
"middle": [
"Hou"
],
"last": "Lei",
"suffix": ""
},
{
"first": "Ravi",
"middle": [],
"last": "Khadiwala",
"suffix": ""
},
{
"first": "Kevin Chen-Chuan",
"middle": [],
"last": "Chang",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of ICDE",
"volume": "",
"issue": "",
"pages": "1273--1276",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rui Li, Kin Hou Lei, Ravi Khadiwala, and Kevin Chen- Chuan Chang. 2012c. Tedas: A twitter-based even- t detection and analysis system. In Proceedings of ICDE, pages 1273-1276, Washington, DC, USA. IEEE Computer Society.",
"links": null
},
"BIBREF14": {
"ref_id": "b14",
"title": "Searching for events in the blogosphere",
"authors": [
{
"first": "Manolis",
"middle": [],
"last": "Platakis",
"suffix": ""
},
{
"first": "Dimitrios",
"middle": [],
"last": "Kotsakos",
"suffix": ""
},
{
"first": "Dimitrios",
"middle": [],
"last": "Gunopulos",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of WWW",
"volume": "",
"issue": "",
"pages": "1225--1226",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Manolis Platakis, Dimitrios Kotsakos, and Dimitrios Gunopulos. 2009. Searching for events in the bl- ogosphere. In Proceedings of WWW, pages 1225- 1226, Madrid, Spain.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "Automatic sub-event detection in emergency management using social media",
"authors": [
{
"first": "Daniela",
"middle": [],
"last": "Pohl",
"suffix": ""
},
{
"first": "Abdelhamid",
"middle": [],
"last": "Bouchachia",
"suffix": ""
},
{
"first": "Hermann",
"middle": [],
"last": "Hellwagner",
"suffix": ""
}
],
"year": 2012,
"venue": "Proceedings of WWW Companion",
"volume": "",
"issue": "",
"pages": "683--686",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Daniela Pohl, Abdelhamid Bouchachia, and Hermann Hellwagner. 2012. Automatic sub-event detection in emergency management using social media. In Proceedings of WWW Companion, pages 683-686, Lyon, France.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Extracting events and event descriptions from twitter",
"authors": [
{
"first": "Ana-Maria",
"middle": [],
"last": "Popescu",
"suffix": ""
},
{
"first": "Marco",
"middle": [],
"last": "Pennacchiotti",
"suffix": ""
},
{
"first": "Deepa",
"middle": [],
"last": "Paranjpe",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of WWW",
"volume": "",
"issue": "",
"pages": "105--106",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ana-Maria Popescu, Marco Pennacchiotti, and Deep- a Paranjpe. 2011. Extracting events and event de- scriptions from twitter. In Proceedings of WWW, pages 105-106, Hyderabad, India.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Earthquake shakes twitter users: real-time event detection by social sensors",
"authors": [
{
"first": "Takeshi",
"middle": [],
"last": "Sakaki",
"suffix": ""
},
{
"first": "Makoto",
"middle": [],
"last": "Okazaki",
"suffix": ""
},
{
"first": "Yutaka",
"middle": [],
"last": "Matsuo",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of WWW",
"volume": "",
"issue": "",
"pages": "851--860",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Takeshi Sakaki, Makoto Okazaki, and Yutaka Matsuo. 2010. Earthquake shakes twitter users: real-time event detection by social sensors. In Proceedings of WWW, pages 851-860, Raleigh, North Carolina, USA.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Feature-rich part-ofspeech tagging with a cyclic dependency network",
"authors": [
{
"first": "Kristina",
"middle": [],
"last": "Toutanova",
"suffix": ""
},
{
"first": "Dan",
"middle": [],
"last": "Klein",
"suffix": ""
},
{
"first": "Christopher",
"middle": [
"D"
],
"last": "Manning",
"suffix": ""
},
{
"first": "Yoram",
"middle": [],
"last": "Singer",
"suffix": ""
}
],
"year": 2003,
"venue": "Proceedings of NAACL",
"volume": "",
"issue": "",
"pages": "173--180",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kristina Toutanova, Dan Klein, Christopher D. Man- ning, and Yoram Singer. 2003. Feature-rich part-of- speech tagging with a cyclic dependency network. In Proceedings of NAACL, pages 173-180, Edmon- ton, Canada.",
"links": null
},
"BIBREF19": {
"ref_id": "b19",
"title": "Location and timeliness of information sources during news events",
"authors": [
{
"first": "Elad",
"middle": [],
"last": "Yom",
"suffix": ""
},
{
"first": "-",
"middle": [],
"last": "Tov",
"suffix": ""
},
{
"first": "Fernando",
"middle": [],
"last": "Diaz",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of SIGIR",
"volume": "",
"issue": "",
"pages": "1105--1106",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Elad Yom-Tov and Fernando Diaz. 2011. Location and timeliness of information sources during news events. In Proceedings of SIGIR, pages 1105-1106, Beijing, China.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Transition-based dependency parsing with rich non-local features",
"authors": [
{
"first": "Yue",
"middle": [],
"last": "Zhang",
"suffix": ""
},
{
"first": "Joakim",
"middle": [],
"last": "Nivre",
"suffix": ""
}
],
"year": 2011,
"venue": "Proceedings of the ACL-HLT",
"volume": "",
"issue": "",
"pages": "188--193",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yue Zhang and Joakim Nivre. 2011. Transition-based dependency parsing with rich non-local features. In Proceedings of the ACL-HLT, pages 188-193, Port- land, Oregon.",
"links": null
}
},
"ref_entries": {
"TABREF1": {
"html": null,
"text": "Experimental Results Using Different Features.",
"num": null,
"type_str": "table",
"content": "
utilized to get system-generated class labels for all