acl acl2011 acl2011-121 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Edward Benson ; Aria Haghighi ; Regina Barzilay
Abstract: We present a novel method for record extraction from social streams such as Twitter. Unlike typical extraction setups, these environments are characterized by short, one sentence messages with heavily colloquial speech. To further complicate matters, individual messages may not express the full relation to be uncovered, as is often assumed in extraction tasks. We develop a graphical model that addresses these problems by learning a latent set of records and a record-message alignment simultaneously; the output of our model is a set of canonical records, the values of which are consistent with aligned messages. We demonstrate that our approach is able to accurately induce event records from Twitter messages, evaluated against events from a local city guide. Our method achieves significant error reduction over baseline methods.1
Reference: text
sentIndex sentText sentNum sentScore
1 edu Abstract We present a novel method for record extraction from social streams such as Twitter. [sent-3, score-0.565]
2 Unlike typical extraction setups, these environments are characterized by short, one sentence messages with heavily colloquial speech. [sent-4, score-0.399]
3 To further complicate matters, individual messages may not express the full relation to be uncovered, as is often assumed in extraction tasks. [sent-5, score-0.369]
4 We develop a graphical model that addresses these problems by learning a latent set of records and a record-message alignment simultaneously; the output of our model is a set of canonical records, the values of which are consistent with aligned messages. [sent-6, score-0.596]
5 We demonstrate that our approach is able to accurately induce event records from Twitter messages, evaluated against events from a local city guide. [sent-7, score-0.626]
6 1 1 Introduction We propose a method for discovering event records from social media feeds such as Twitter. [sent-9, score-0.631]
7 Social media messages are often short, make heavy use of colloquial language, and require situational context for interpretation (see examples in Figure 1). [sent-13, score-0.4]
8 Not all properties of an event may be expressed in a single message, and the mapping between messages and canonical event records is not obvious. [sent-14, score-1.019]
9 Despite these challenges, this data exhibits an important property that makes learning amenable: the multitude of messages referencing the same event. [sent-26, score-0.283]
10 Our goal is to induce a comprehensive set of event records given a seed set of example records, such as a city event calendar table. [sent-27, score-0.74]
11 At the message level, the model relies on a conditional random field component to extract field values such ProceedinPgosrt olafn thde, 4 O9rtehg Aonn,n Juuanle M 1e9e-2tin4g, 2 o0f1 t1h. [sent-31, score-0.492]
12 Ac s2s0o1ci1a Atiosnso fcoirat Cioonm foprut Caotimonpaulta Lti nognuails Lti cnsg,u piasgteics 389–398, as location of the event and artist name. [sent-33, score-0.244]
13 We bias local decisions made by the CRF to be consistent with canonical record values, thereby facilitating consistency within an event cluster. [sent-34, score-0.696]
14 A seed set of example records constitutes our only source of supervision; we do not observe alignment between these seed records and individual messages, nor any message-level field annotation. [sent-37, score-0.921]
15 The output of our model consists of an event-based clustering of messages, where each cluster is represented by a single multi-field record with a canonical value chosen for each field. [sent-38, score-0.53]
16 We apply our technique to construct entertainment event records for the city calendar section of NYC . [sent-39, score-0.603]
17 Identifying messages that re390 fer to the same event is a large part of our challenge. [sent-53, score-0.42]
18 (2009b) induces the mapping automatically by bootstrapping from sentences that directly match record entries. [sent-60, score-0.431]
19 The key challenge for our method is modeling message to record alignment which is not an issue in the previous set up. [sent-67, score-0.688]
20 To our knowledge no work has yet addressed record extraction from this growing corpus. [sent-72, score-0.481]
21 Message (x): Each message x is a single posting to Twitter. [sent-75, score-0.23]
22 Record (R): A record is a representation of the canonical properties of an event. [sent-78, score-0.496]
23 We use Ri to de- note the ith record and Ri‘ to denote the value of the ‘th property of that record. [sent-79, score-0.431]
24 In our experiments, each record Ri is a tuple hR1i , Ri2i which represents that Figure 2: The key variables of our model. [sent-80, score-0.47]
25 A collection of K latent records Rk, each consisting of a set of L properties. [sent-81, score-0.434]
26 ” Each tweet xi is associated with a labeling over tokens yi and is aligned to a record via the Ai variable. [sent-83, score-0.619]
27 Message Labels (y): We assume that each message has a sequence labeling, where the labels consist of the record fields (e. [sent-94, score-0.714]
28 Each token xj in a message has an associated label yj. [sent-97, score-0.428]
29 Message to Record Alignment (A): We assume that each message is aligned to some record such that the event described in the message is the one represented by that record. [sent-99, score-1.06]
30 Each message xi is associated with an alignment variable Ai that takes a value in {1, . [sent-100, score-0.257]
31 Multiple messages can atn odf do align to the same record. [sent-106, score-0.283]
32 As discussed in Section 4, our model will encourage tokens associated with message labels to be “similar” to corresponding aligned record values. [sent-107, score-0.758]
33 , RK‘ of record values for a particular domain field ‘. [sent-118, score-0.569]
34 , 2001), where the potential over a message label sequence decomposes YiφSEQXiφUNQR R k +k−1 (acrtohs rfieceolrd s)RkφPOPXYAi k thRrkecorRdk+1φCONXYAi Figure 3: Factor graph representation of our model. [sent-123, score-0.308]
35 Many of the features characterize how likely a given token label, such as ARTIST, is for a given position in the message sequence conditioning arbitrarily on message text context. [sent-130, score-0.57]
36 As a result some events may have only a few referent messages while other more popular events may have thousands or more. [sent-136, score-0.395]
37 In such a circumstance, the messages for a popular event may collect to form multiple identical record clusters. [sent-137, score-0.851]
38 : xxx, XXX, Xxx, or other 3These are just features, not a filter; we are free to extract any artist or venue regardless of their inclusion in this list. [sent-140, score-0.238]
39 392 fix the number of records learned, such behavior inhibits the discovery of less talked-about events. [sent-141, score-0.397]
40 Instead, we would rather have just two records: one with two aligned messages and another with thousands. [sent-142, score-0.315]
41 This factor fac- torizes over pairs of records: φUNQ(R‘) = Y φUNQ(R‘k, Rk‘0) kY6=k0 where Rk‘ and Rk‘0 are the values of field ‘ for two records Rk and Rk0. [sent-148, score-0.642]
42 Because speech on Twitter is colloquial, we would like these clusters to be amenable to many variations of the canonical record properties that are ultimately learned. [sent-152, score-0.496]
43 The φPOP factor accomplishes this by representing a lenient compatibility score between a message x, its labels y, and some candidate value v for a record field (e. [sent-153, score-0.868]
44 This factor decomposes over tokens, and we align each token xj with the best matching token vk in v (e. [sent-156, score-0.504]
45 The token level sum is scaled by the length of the record value being matched to avoid a preference for long field values. [sent-159, score-0.643]
46 φPOP(x, y, RA‘ = v) = XjmkaxφPOP(xj,y|vj|,RA‘= vk) This token-level component may be thought of as a compatibility score between the labeled token xj and the record field assignment RA‘ = v. [sent-160, score-0.777]
47 Given that token xj aligns with the token vk, the token-level component returns the sum of three parts, subject to the constraint that yj = ‘: • IDF(xj)I[xj = vk], an equality indicator betIwDeFen( xtokens xj and vk, scaled by the inverse document frequency of xj • αIDF(xj) ? [sent-161, score-0.585]
48 4 Record Consistency Factor While the uniqueness factor discourages a flood of messages for a single event from clustering into multiple event records, we also wish to discourage messages from multiple events from clustering into the same record. [sent-168, score-1.164]
49 When such a situation occurs, the model may either resolve it by changing inconsistent token labelings to the NONE label or by reassigning some of the messages to a new cluster. [sent-169, score-0.398]
50 We encourage the latter solution with a record consistency factor φCON. [sent-170, score-0.626]
51 The record consistency factor is an indicator function on the field values of a record being present and labeled correctly in a message. [sent-171, score-1.194]
52 While the popularity factor encourages agreement on a per-label basis, this factor influences the joint behavior of message labels to agree with the aligned record. [sent-172, score-0.526]
53 5 Inference Our goal is to predict a set of records R. [sent-177, score-0.397]
54 q(y) update: The q(y) update for a single message yields an implicit expression in terms of pairwise cliques in y. [sent-190, score-0.332]
55 Inference consists of performing updates for each of the three kinds of latent variables: message labels (y), record alignments (A), and record field values (R‘). [sent-195, score-1.301]
56 All are relatively cheap to compute except for the record field update q(Rk‘) which requires looping potentially over all messages. [sent-196, score-0.641]
57 Therefore computing the q(y) update amounts to re-running forward backwards on the message where there is an expected potential term which involves the belief over other variables. [sent-199, score-0.336]
58 Note that the popularity and consensus potentials (φPOP and φCON) decompose over individual message tokens so this can be tractably computed. [sent-200, score-0.389]
59 q(A) update: The update for individual record alignment reduces to being log-proportional to the expected popularity and consensus potentials. [sent-201, score-0.61]
60 q(Rk‘) update: The update for the record field 394 distribution is the most complex factor of the three. [sent-202, score-0.715]
61 It requires computing expected similarity with other record field values (the φUNQ potential) and looping over all messages to accumulate a contribution from each, weighted by the probability that it is aligned to the target record. [sent-203, score-0.917]
62 1 Initializing Factors Since a uniform initialization of all factors is a saddle-point of the objective, we opt to initialize the q(y) factors with the marginals obtained using just the CRF parameters, accomplished by running forwards-backwards on all messages using only the φSEQ potentials. [sent-205, score-0.345]
63 Because each term in φPOP and φCON includes an indicator function based on a token match between a field-value and a message, knowing the possible values v of each Rk‘ enables us to precompute the combinations of (x, ‘, v) for which nonzero factor values are possible. [sent-213, score-0.302]
64 The messages have an average length of 18 tokens, and the corpus vocabulary comprises 468,000 unique words6. [sent-218, score-0.283]
65 We obtain labeled gold records using data scraped from the NYC . [sent-219, score-0.535]
66 Each gold record had two fields of interest: ARTIST and VENUE. [sent-221, score-0.508]
67 Preprocessing Only a small fraction of Twitter messages are relevant to the target extraction task. [sent-223, score-0.333]
68 , 2010) to predict whether a message mentions a concert event. [sent-229, score-0.23]
69 While the two-stage filtering does not fully eliminate noise in the input stream, it greatly reduces the presence of irrelevant messages to a manageable 5,800 messages without filtering too many ‘signal’ tweets. [sent-234, score-0.566]
70 We also filter our gold record set to include only records in which each field value occurs at least once somewhere in the corpus, as these are the records which are possible to learn given the input. [sent-235, score-1.403]
71 Training The first weekend of data (2,184 messages and 11records after preprocessing) is used for training. [sent-237, score-0.332]
72 While it is possible to train these parameters via direct annotation of messages with label sequences, we opted instead to use a simple approach where message tokens from the training weekend are labeled via their intersection with gold records, often called “distant supervision” (Mintz et al. [sent-239, score-0.674]
73 Concretely, we automatically label message tokens in the training corpus with either the ARTIST or VENUE label if they belonged to a sequence that matched a gold record field, and with NONE otherwise. [sent-241, score-0.773]
74 This is the only use that is made of the gold records throughout training. [sent-242, score-0.445]
75 Testing The two weekends of data used for testing totaled 3,662 tweets after preprocessing and 3 1 gold records for evaluation. [sent-244, score-0.554]
76 Our model assumes a fixed number of records K = 130. [sent-246, score-0.397]
77 The horizontal axis is the number of records kept from the ranked model output, as a multiple of the number of golds. [sent-248, score-0.397]
78 The CRF lines terminate because of low record yield. [sent-249, score-0.456]
79 This ranking function is intended to push garbage collection records to the bottom of the list. [sent-251, score-0.426]
80 The baselines label each message and then extract one record for each combination of labeled phrases. [sent-255, score-0.685]
81 Our List Baseline labels messages by finding string overlaps against a list of musical artists and venues scraped from web data (the same lists used as features in our CRF component). [sent-257, score-0.378]
82 The Low Threshold Baseline generates all possible records from labelings with a token-level likelihood greater than λ = 0. [sent-259, score-0.426]
83 The output of these baselines is a set of records ranked by the number of votes cast for each, and we perform our evaluation against the top k of these records. [sent-261, score-0.397]
84 7 Evaluation The evaluation of record construction is challenging because many induced music events discussed 396 in Twitter messages are not in our gold data set; our gold records are precise but incomplete. [sent-262, score-1.263]
85 Both evaluations are performed using hard zero-one loss at record level. [sent-264, score-0.431]
86 Recall We evaluate recall, shown in Figure 5, against the gold event records for each weekend. [sent-266, score-0.582]
87 We perform our evaluation by taking the top k records induced, performing a stable marriage matching against the gold records, and then evaluating the resulting matched pairs. [sent-268, score-0.478]
88 Precision To evaluate precision we assembled a list of the distinct records produced by all models and then manually determined if each record was correct. [sent-273, score-0.878]
89 We then used this aggregate list of correct records to measure precision for each individual model, shown in Figure 6. [sent-275, score-0.423]
90 The downward trend in precision that can be seen in Figure 6 is the effect of our ranking algorithm, which attempts to push garbage collection records towards the bottom of the record list. [sent-279, score-0.883]
91 The CRF and hard-constrained consensus lines terminate because of low record yield. [sent-282, score-0.481]
92 Our model’s soft string comparison-based clustering can be seen at work when our model uncovers records that would have been impossible without this approach. [sent-293, score-0.431]
93 Terminal Five → 397 Terminal 5) even when no message about the event spells the venue correctly. [sent-296, score-0.498]
94 Still, the clustering can introduce errors by combining messages that provide orthogonal field contributions yet have overlapping tokens (thus escaping the penalty of the consistency factor). [sent-297, score-0.52]
95 An example of two messages participating in this scenario is shown below; the shared term “holiday” in the second message gets relabeled as ARTIST: CPlosm tuen ceh einck to ou TtV the Gu hiodlieda Nye ctwheoerkrAVretnisutepTaOrkNsiIdGeH isT bu artst 8in pgm. [sent-298, score-0.513]
96 Because short messages are unlikely to express high arity relations completely, tying extraction and clustering seems an intuitive solution. [sent-304, score-0.367]
97 In such a scenario, the record consistency constraints imposed by our model would have to be relaxed, perhaps examining pairwise argument consistency instead. [sent-305, score-0.557]
98 8 Conclusion We presented a novel model for record extraction from social media streams such as Twitter. [sent-306, score-0.616]
99 Our model operates on a noisy feed of data and extracts canonical records of events by aggregating information across multiple messages. [sent-307, score-0.518]
100 Despite the noise of irrelevant messages and the relatively colloquial nature of message language, we are able to extract records with relatively high accuracy. [sent-308, score-0.976]
wordName wordTfidf (topN-words)
[('record', 0.431), ('records', 0.397), ('messages', 0.283), ('pop', 0.24), ('message', 0.23), ('rk', 0.211), ('unq', 0.181), ('con', 0.158), ('event', 0.137), ('seq', 0.133), ('venue', 0.131), ('xj', 0.112), ('artist', 0.107), ('factor', 0.107), ('twitter', 0.106), ('field', 0.1), ('lnh', 0.099), ('crf', 0.097), ('uniqueness', 0.093), ('vk', 0.088), ('token', 0.086), ('update', 0.077), ('mintz', 0.071), ('variational', 0.07), ('colloquial', 0.066), ('scraped', 0.066), ('weekends', 0.066), ('canonical', 0.065), ('yi', 0.065), ('consistency', 0.063), ('yao', 0.061), ('events', 0.056), ('media', 0.051), ('labeling', 0.051), ('extraction', 0.05), ('popularity', 0.05), ('lnq', 0.049), ('rz', 0.049), ('weekend', 0.049), ('gold', 0.048), ('social', 0.046), ('supervision', 0.044), ('potentials', 0.044), ('nyc', 0.043), ('tweets', 0.043), ('distant', 0.041), ('tokens', 0.04), ('variables', 0.039), ('stream', 0.039), ('ln', 0.038), ('values', 0.038), ('xxx', 0.038), ('streams', 0.038), ('latent', 0.037), ('city', 0.036), ('relation', 0.036), ('mann', 0.035), ('updates', 0.034), ('clustering', 0.034), ('ai', 0.034), ('calendar', 0.033), ('fsim', 0.033), ('holiday', 0.033), ('looping', 0.033), ('marriage', 0.033), ('nlnh', 0.033), ('precompute', 0.033), ('screw', 0.033), ('thresh', 0.033), ('aligned', 0.032), ('factors', 0.031), ('ritter', 0.031), ('names', 0.031), ('io', 0.03), ('filter', 0.03), ('garbage', 0.029), ('artists', 0.029), ('labelings', 0.029), ('yij', 0.029), ('potential', 0.029), ('fields', 0.029), ('alignment', 0.027), ('inference', 0.027), ('ra', 0.027), ('dave', 0.027), ('irving', 0.027), ('constraint', 0.027), ('scaled', 0.026), ('precision', 0.026), ('idf', 0.026), ('consensus', 0.025), ('cliques', 0.025), ('decomposes', 0.025), ('terminate', 0.025), ('encourage', 0.025), ('component', 0.024), ('sequence', 0.024), ('labeled', 0.024), ('yarowsky', 0.024), ('distinct', 0.024)]
simIndex simValue paperId paperTitle
same-paper 1 0.99999958 121 acl-2011-Event Discovery in Social Media Feeds
Author: Edward Benson ; Aria Haghighi ; Regina Barzilay
Abstract: We present a novel method for record extraction from social streams such as Twitter. Unlike typical extraction setups, these environments are characterized by short, one sentence messages with heavily colloquial speech. To further complicate matters, individual messages may not express the full relation to be uncovered, as is often assumed in extraction tasks. We develop a graphical model that addresses these problems by learning a latent set of records and a record-message alignment simultaneously; the output of our model is a set of canonical records, the values of which are consistent with aligned messages. We demonstrate that our approach is able to accurately induce event records from Twitter messages, evaluated against events from a local city guide. Our method achieves significant error reduction over baseline methods.1
2 0.14822099 36 acl-2011-An Efficient Indexer for Large N-Gram Corpora
Author: Hakan Ceylan ; Rada Mihalcea
Abstract: We introduce a new publicly available tool that implements efficient indexing and retrieval of large N-gram datasets, such as the Web1T 5-gram corpus. Our tool indexes the entire Web1T dataset with an index size of only 100 MB and performs a retrieval of any N-gram with a single disk access. With an increased index size of 420 MB and duplicate data, it also allows users to issue wild card queries provided that the wild cards in the query are contiguous. Furthermore, we also implement some of the smoothing algorithms that are designed specifically for large datasets and are shown to yield better language models than the traditional ones on the Web1T 5gram corpus (Yuret, 2008). We demonstrate the effectiveness of our tool and the smoothing algorithms on the English Lexical Substi- tution task by a simple implementation that gives considerable improvement over a basic language model.
3 0.14455567 328 acl-2011-Using Cross-Entity Inference to Improve Event Extraction
Author: Yu Hong ; Jianfeng Zhang ; Bin Ma ; Jianmin Yao ; Guodong Zhou ; Qiaoming Zhu
Abstract: Event extraction is the task of detecting certain specified types of events that are mentioned in the source language data. The state-of-the-art research on the task is transductive inference (e.g. cross-event inference). In this paper, we propose a new method of event extraction by well using cross-entity inference. In contrast to previous inference methods, we regard entitytype consistency as key feature to predict event mentions. We adopt this inference method to improve the traditional sentence-level event extraction system. Experiments show that we can get 8.6% gain in trigger (event) identification, and more than 11.8% gain for argument (role) classification in ACE event extraction. 1
4 0.14110433 65 acl-2011-Can Document Selection Help Semi-supervised Learning? A Case Study On Event Extraction
Author: Shasha Liao ; Ralph Grishman
Abstract: Annotating training data for event extraction is tedious and labor-intensive. Most current event extraction tasks rely on hundreds of annotated documents, but this is often not enough. In this paper, we present a novel self-training strategy, which uses Information Retrieval (IR) to collect a cluster of related documents as the resource for bootstrapping. Also, based on the particular characteristics of this corpus, global inference is applied to provide more confident and informative data selection. We compare this approach to self-training on a normal newswire corpus and show that IR can provide a better corpus for bootstrapping and that global inference can further improve instance selection. We obtain gains of 1.7% in trigger labeling and 2.3% in role labeling through IR and an additional 1.1% in trigger labeling and 1.3% in role labeling by applying global inference. 1
5 0.13058387 244 acl-2011-Peeling Back the Layers: Detecting Event Role Fillers in Secondary Contexts
Author: Ruihong Huang ; Ellen Riloff
Abstract: The goal of our research is to improve event extraction by learning to identify secondary role filler contexts in the absence of event keywords. We propose a multilayered event extraction architecture that progressively “zooms in” on relevant information. Our extraction model includes a document genre classifier to recognize event narratives, two types of sentence classifiers, and noun phrase classifiers to extract role fillers. These modules are organized as a pipeline to gradually zero in on event-related information. We present results on the MUC-4 event extraction data set and show that this model performs better than previous systems.
6 0.12633736 122 acl-2011-Event Extraction as Dependency Parsing
7 0.11933298 242 acl-2011-Part-of-Speech Tagging for Twitter: Annotation, Features, and Experiments
8 0.11862677 208 acl-2011-Lexical Normalisation of Short Text Messages: Makn Sens a #twitter
9 0.10714084 190 acl-2011-Knowledge-Based Weak Supervision for Information Extraction of Overlapping Relations
10 0.10470065 172 acl-2011-Insertion, Deletion, or Substitution? Normalizing Text Messages without Pre-categorization nor Supervision
11 0.10190593 170 acl-2011-In-domain Relation Discovery with Meta-constraints via Posterior Regularization
12 0.098742977 194 acl-2011-Language Use: What can it tell us?
13 0.094281659 177 acl-2011-Interactive Group Suggesting for Twitter
14 0.083992422 261 acl-2011-Recognizing Named Entities in Tweets
15 0.083343655 293 acl-2011-Template-Based Information Extraction without the Templates
16 0.081516504 226 acl-2011-Multi-Modal Annotation of Quest Games in Second Life
17 0.079732478 160 acl-2011-Identifying Sarcasm in Twitter: A Closer Look
18 0.074286267 218 acl-2011-MemeTube: A Sentiment-based Audiovisual System for Analyzing and Displaying Microblog Messages
19 0.074146792 325 acl-2011-Unsupervised Word Alignment with Arbitrary Features
20 0.073647007 286 acl-2011-Social Network Extraction from Texts: A Thesis Proposal
topicId topicWeight
[(0, 0.169), (1, 0.08), (2, -0.084), (3, 0.004), (4, 0.111), (5, 0.071), (6, -0.018), (7, -0.086), (8, 0.056), (9, 0.085), (10, -0.07), (11, 0.135), (12, 0.103), (13, 0.037), (14, -0.087), (15, -0.034), (16, -0.003), (17, -0.024), (18, -0.03), (19, -0.003), (20, 0.026), (21, -0.015), (22, 0.003), (23, -0.012), (24, -0.008), (25, -0.042), (26, 0.018), (27, -0.005), (28, 0.047), (29, -0.037), (30, -0.015), (31, -0.011), (32, 0.027), (33, 0.061), (34, -0.009), (35, 0.027), (36, -0.026), (37, 0.054), (38, -0.024), (39, -0.03), (40, 0.013), (41, 0.004), (42, 0.083), (43, -0.056), (44, -0.058), (45, -0.046), (46, 0.009), (47, 0.014), (48, 0.075), (49, -0.028)]
simIndex simValue paperId paperTitle
same-paper 1 0.94577348 121 acl-2011-Event Discovery in Social Media Feeds
Author: Edward Benson ; Aria Haghighi ; Regina Barzilay
Abstract: We present a novel method for record extraction from social streams such as Twitter. Unlike typical extraction setups, these environments are characterized by short, one sentence messages with heavily colloquial speech. To further complicate matters, individual messages may not express the full relation to be uncovered, as is often assumed in extraction tasks. We develop a graphical model that addresses these problems by learning a latent set of records and a record-message alignment simultaneously; the output of our model is a set of canonical records, the values of which are consistent with aligned messages. We demonstrate that our approach is able to accurately induce event records from Twitter messages, evaluated against events from a local city guide. Our method achieves significant error reduction over baseline methods.1
2 0.58834738 242 acl-2011-Part-of-Speech Tagging for Twitter: Annotation, Features, and Experiments
Author: Kevin Gimpel ; Nathan Schneider ; Brendan O'Connor ; Dipanjan Das ; Daniel Mills ; Jacob Eisenstein ; Michael Heilman ; Dani Yogatama ; Jeffrey Flanigan ; Noah A. Smith
Abstract: We address the problem of part-of-speech tagging for English data from the popular microblogging service Twitter. We develop a tagset, annotate data, develop features, and report tagging results nearing 90% accuracy. The data and tools have been made available to the research community with the goal of enabling richer text analysis of Twitter and related social media data sets.
3 0.57853973 244 acl-2011-Peeling Back the Layers: Detecting Event Role Fillers in Secondary Contexts
Author: Ruihong Huang ; Ellen Riloff
Abstract: The goal of our research is to improve event extraction by learning to identify secondary role filler contexts in the absence of event keywords. We propose a multilayered event extraction architecture that progressively “zooms in” on relevant information. Our extraction model includes a document genre classifier to recognize event narratives, two types of sentence classifiers, and noun phrase classifiers to extract role fillers. These modules are organized as a pipeline to gradually zero in on event-related information. We present results on the MUC-4 event extraction data set and show that this model performs better than previous systems.
4 0.57738036 122 acl-2011-Event Extraction as Dependency Parsing
Author: David McClosky ; Mihai Surdeanu ; Christopher Manning
Abstract: Nested event structures are a common occurrence in both open domain and domain specific extraction tasks, e.g., a “crime” event can cause a “investigation” event, which can lead to an “arrest” event. However, most current approaches address event extraction with highly local models that extract each event and argument independently. We propose a simple approach for the extraction of such structures by taking the tree of event-argument relations and using it directly as the representation in a reranking dependency parser. This provides a simple framework that captures global properties of both nested and flat event structures. We explore a rich feature space that models both the events to be parsed and context from the original supporting text. Our approach obtains competitive results in the extraction of biomedical events from the BioNLP’09 shared task with a F1 score of 53.5% in development and 48.6% in testing.
Author: Fei Liu ; Fuliang Weng ; Bingqing Wang ; Yang Liu
Abstract: Most text message normalization approaches are based on supervised learning and rely on human labeled training data. In addition, the nonstandard words are often categorized into different types and specific models are designed to tackle each type. In this paper, we propose a unified letter transformation approach that requires neither pre-categorization nor human supervision. Our approach models the generation process from the dictionary words to nonstandard tokens under a sequence labeling framework, where each letter in the dictionary word can be retained, removed, or substituted by other letters/digits. To avoid the expensive and time consuming hand labeling process, we automatically collected a large set of noisy training pairs using a novel webbased approach and performed character-level . alignment for model training. Experiments on both Twitter and SMS messages show that our system significantly outperformed the stateof-the-art deletion-based abbreviation system and the jazzy spell checker (absolute accuracy gain of 21.69% and 18. 16% over jazzy spell checker on the two test sets respectively).
6 0.57569122 65 acl-2011-Can Document Selection Help Semi-supervised Learning? A Case Study On Event Extraction
7 0.56862932 226 acl-2011-Multi-Modal Annotation of Quest Games in Second Life
8 0.56254262 328 acl-2011-Using Cross-Entity Inference to Improve Event Extraction
9 0.53501898 208 acl-2011-Lexical Normalisation of Short Text Messages: Makn Sens a #twitter
10 0.52997625 190 acl-2011-Knowledge-Based Weak Supervision for Information Extraction of Overlapping Relations
11 0.52908379 291 acl-2011-SystemT: A Declarative Information Extraction System
12 0.52885377 160 acl-2011-Identifying Sarcasm in Twitter: A Closer Look
13 0.52114487 286 acl-2011-Social Network Extraction from Texts: A Thesis Proposal
14 0.5162679 293 acl-2011-Template-Based Information Extraction without the Templates
15 0.50451922 261 acl-2011-Recognizing Named Entities in Tweets
16 0.47699788 177 acl-2011-Interactive Group Suggesting for Twitter
17 0.45621461 186 acl-2011-Joint Training of Dependency Parsing Filters through Latent Support Vector Machines
18 0.45613694 97 acl-2011-Discovering Sociolinguistic Associations with Structured Sparsity
19 0.44533357 170 acl-2011-In-domain Relation Discovery with Meta-constraints via Posterior Regularization
20 0.44091532 194 acl-2011-Language Use: What can it tell us?
topicId topicWeight
[(5, 0.031), (9, 0.013), (17, 0.042), (18, 0.214), (26, 0.023), (37, 0.082), (39, 0.047), (41, 0.071), (52, 0.013), (55, 0.052), (59, 0.058), (62, 0.02), (72, 0.042), (91, 0.049), (96, 0.111), (97, 0.01)]
simIndex simValue paperId paperTitle
same-paper 1 0.81748295 121 acl-2011-Event Discovery in Social Media Feeds
Author: Edward Benson ; Aria Haghighi ; Regina Barzilay
Abstract: We present a novel method for record extraction from social streams such as Twitter. Unlike typical extraction setups, these environments are characterized by short, one sentence messages with heavily colloquial speech. To further complicate matters, individual messages may not express the full relation to be uncovered, as is often assumed in extraction tasks. We develop a graphical model that addresses these problems by learning a latent set of records and a record-message alignment simultaneously; the output of our model is a set of canonical records, the values of which are consistent with aligned messages. We demonstrate that our approach is able to accurately induce event records from Twitter messages, evaluated against events from a local city guide. Our method achieves significant error reduction over baseline methods.1
2 0.7289232 162 acl-2011-Identifying the Semantic Orientation of Foreign Words
Author: Ahmed Hassan ; Amjad AbuJbara ; Rahul Jha ; Dragomir Radev
Abstract: We present a method for identifying the positive or negative semantic orientation of foreign words. Identifying the semantic orientation of words has numerous applications in the areas of text classification, analysis of product review, analysis of responses to surveys, and mining online discussions. Identifying the semantic orientation of English words has been extensively studied in literature. Most of this work assumes the existence of resources (e.g. Wordnet, seeds, etc) that do not exist in foreign languages. In this work, we describe a method based on constructing a multilingual network connecting English and foreign words. We use this network to identify the semantic orientation of foreign words based on connection between words in the same language as well as multilingual connections. The method is experimentally tested using a manually labeled set of positive and negative words and has shown very promising results.
3 0.71648157 182 acl-2011-Joint Annotation of Search Queries
Author: Michael Bendersky ; W. Bruce Croft ; David A. Smith
Abstract: W. Bruce Croft Dept. of Computer Science University of Massachusetts Amherst, MA cro ft @ c s .uma s s .edu David A. Smith Dept. of Computer Science University of Massachusetts Amherst, MA dasmith@ c s .umas s .edu articles or web pages). As previous research shows, these differences severely limit the applicability of Marking up search queries with linguistic annotations such as part-of-speech tags, capitalization, and segmentation, is an impor- tant part of query processing and understanding in information retrieval systems. Due to their brevity and idiosyncratic structure, search queries pose a challenge to existing NLP tools. To address this challenge, we propose a probabilistic approach for performing joint query annotation. First, we derive a robust set of unsupervised independent annotations, using queries and pseudo-relevance feedback. Then, we stack additional classifiers on the independent annotations, and exploit the dependencies between them to further improve the accuracy, even with a very limited amount of available training data. We evaluate our method using a range of queries extracted from a web search log. Experimental results verify the effectiveness of our approach for both short keyword queries, and verbose natural language queries.
4 0.69512278 327 acl-2011-Using Bilingual Parallel Corpora for Cross-Lingual Textual Entailment
Author: Yashar Mehdad ; Matteo Negri ; Marcello Federico
Abstract: This paper explores the use of bilingual parallel corpora as a source of lexical knowledge for cross-lingual textual entailment. We claim that, in spite of the inherent difficulties of the task, phrase tables extracted from parallel data allow to capture both lexical relations between single words, and contextual information useful for inference. We experiment with a phrasal matching method in order to: i) build a system portable across languages, and ii) evaluate the contribution of lexical knowledge in isolation, without interaction with other inference mechanisms. Results achieved on an English-Spanish corpus obtained from the RTE3 dataset support our claim, with an overall accuracy above average scores reported by RTE participants on monolingual data. Finally, we show that using parallel corpora to extract paraphrase tables reveals their potential also in the monolingual setting, improving the results achieved with other sources of lexical knowledge.
5 0.64351022 324 acl-2011-Unsupervised Semantic Role Induction via Split-Merge Clustering
Author: Joel Lang ; Mirella Lapata
Abstract: In this paper we describe an unsupervised method for semantic role induction which holds promise for relieving the data acquisition bottleneck associated with supervised role labelers. We present an algorithm that iteratively splits and merges clusters representing semantic roles, thereby leading from an initial clustering to a final clustering of better quality. The method is simple, surprisingly effective, and allows to integrate linguistic knowledge transparently. By combining role induction with a rule-based component for argument identification we obtain an unsupervised end-to-end semantic role labeling system. Evaluation on the CoNLL 2008 benchmark dataset demonstrates that our method outperforms competitive unsupervised approaches by a wide margin.
6 0.64073771 119 acl-2011-Evaluating the Impact of Coder Errors on Active Learning
7 0.64032876 170 acl-2011-In-domain Relation Discovery with Meta-constraints via Posterior Regularization
8 0.63977194 126 acl-2011-Exploiting Syntactico-Semantic Structures for Relation Extraction
9 0.63619018 145 acl-2011-Good Seed Makes a Good Crop: Accelerating Active Learning Using Language Modeling
10 0.63439727 65 acl-2011-Can Document Selection Help Semi-supervised Learning? A Case Study On Event Extraction
11 0.6336993 190 acl-2011-Knowledge-Based Weak Supervision for Information Extraction of Overlapping Relations
12 0.63244456 58 acl-2011-Beam-Width Prediction for Efficient Context-Free Parsing
13 0.63149595 269 acl-2011-Scaling up Automatic Cross-Lingual Semantic Role Annotation
14 0.63070387 164 acl-2011-Improving Arabic Dependency Parsing with Form-based and Functional Morphological Features
15 0.63013518 209 acl-2011-Lexically-Triggered Hidden Markov Models for Clinical Document Coding
16 0.62995338 86 acl-2011-Coreference for Learning to Extract Relations: Yes Virginia, Coreference Matters
17 0.62840003 36 acl-2011-An Efficient Indexer for Large N-Gram Corpora
18 0.62812436 3 acl-2011-A Bayesian Model for Unsupervised Semantic Parsing
19 0.62621313 196 acl-2011-Large-Scale Cross-Document Coreference Using Distributed Inference and Hierarchical Models
20 0.62611413 246 acl-2011-Piggyback: Using Search Engines for Robust Cross-Domain Named Entity Recognition