acl acl2013 acl2013-373 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Shane Bergsma ; Benjamin Van Durme
Abstract: We describe a novel approach for automatically predicting the hidden demographic properties of social media users. Building on prior work in common-sense knowledge acquisition from third-person text, we first learn the distinguishing attributes of certain classes of people. For example, we learn that people in the Female class tend to have maiden names and engagement rings. We then show that this knowledge can be used in the analysis of first-person communication; knowledge of distinguishing attributes allows us to both classify users and to bootstrap new training examples. Our novel approach enables substantial improvements on the widelystudied task of user gender prediction, ob- taining a 20% relative error reduction over the current state-of-the-art.
Reference: text
sentIndex sentText sentNum sentScore
1 Building on prior work in common-sense knowledge acquisition from third-person text, we first learn the distinguishing attributes of certain classes of people. [sent-2, score-0.576]
2 For example, we learn that people in the Female class tend to have maiden names and engagement rings. [sent-3, score-0.159]
3 We then show that this knowledge can be used in the analysis of first-person communication; knowledge of distinguishing attributes allows us to both classify users and to bootstrap new training examples. [sent-4, score-0.638]
4 Our novel approach enables substantial improvements on the widelystudied task of user gender prediction, ob- taining a 20% relative error reduction over the current state-of-the-art. [sent-5, score-0.357]
5 1 Introduction There has been growing interest in characterizing social media users based on the content they generate; that is, automatically labeling users with demographic categories such as age and gender (Burger and Henderson, 2006; Schler et al. [sent-6, score-0.643]
6 Automatic user characterization has applications in targeted advertising and personalization, and could also lead to finergrained assessment of public opinion (O’Connor et al. [sent-10, score-0.264]
7 We present an algorithm that improves user characterization by collecting and exploiting such common-sense knowledge. [sent-20, score-0.264]
8 Our work is inspired by algorithms that processes large text corpora in order to discover the attributes of semantic classes, e. [sent-21, score-0.432]
9 We learn the distinguishing attributes of different demographic groups (Section 3), and then automatically assign users to these groups whenever they refer to a distinguishing attribute in their writings (Section 4). [sent-27, score-0.997]
10 We validate our approach by advancing the state-of-the-art on the most well-studied user classification task: predicting user gender (Section 5). [sent-29, score-0.548]
11 Our bootstrapped system, trained purely from automatically-annotated Twitter data, significantly reduces error over a state-of-the-art system trained on thousands of gold-standard training examples. [sent-30, score-0.22]
12 2 Supervised User Characterization The current state-of-the-art in user characterization is to use supervised classifiers trained on annotated data. [sent-31, score-0.342]
13 User-Profile Features: Usr Some researchers have explored features for user-profile metainformation in addition to user content. [sent-40, score-0.174]
14 Using a combination of content and username features “represents a use case common to many different social media sites, such as chat rooms and news article comment streams” (Burger et al. [sent-48, score-0.296]
15 We refer to features derived from a username as Usr features in our experiments. [sent-50, score-0.19]
16 3 Learning Class Attributes We aim to improve the automated classification of users into various demographic categories by learning and applying the distinguishing attributes of those categories, e. [sent-51, score-0.774]
17 In this research, the objective is to discover various attributes or parts of classes of entities. [sent-55, score-0.469]
18 For example, Berland and Charniak (1999) learn that the class car has parts such as headlight, windshield, dashboard, etc. [sent-56, score-0.15]
19 Berland and Charniak extract these attributes by mining a corpus for fillers of patterns such as ‘car’s X’ or ‘X of a car’ . [sent-57, score-0.527]
20 Note their patterns explicitly include the class itself (car). [sent-58, score-0.161]
21 For example, Pas ¸ca and Van Durme (2007) learn the attributes of the class car via patterns involving instances of cars, e. [sent-62, score-0.679]
22 For the gender task that we study in our experiments, we acquire class instances by filtering the dataset of nouns and their genders created by Bergsma and Lin (2006). [sent-69, score-0.362]
23 We extract prevalent common nouns for males and females by selecting only those nouns that (a) occur more than 200 times in the dataset, (b) mostly occur with male or female pronouns, and (c) occur as lower-case more often than upper-case in a web-scale N-gram corpus (Lin et al. [sent-71, score-0.376]
24 Since the gender data is noisy, we also quickly pruned by hand any instances that were malformed or obviously incorrectly assigned by our automatic process. [sent-76, score-0.259]
25 In attribute extraction, typically one must choose between the precise results of rich patterns (involving punctuation and parts-of-speech) applied to small corpora (Berland and Charniak, 1999) and the high-coverage results of superficial patterns applied to web-scale data, e. [sent-98, score-0.286]
26 Prior work has mostly focused on finding “relevant” attributes (Alfonseca et al. [sent-106, score-0.432]
27 A leg is a relevant and correct part of both a male and a female (and many other living and inanimate objects), but it does not help us distinguish males from females in social media. [sent-108, score-0.451]
28 We therefore rank our attributes for each class by their strength of association with instances of that specific class. [sent-109, score-0.574]
29 For each class, we rank the attributes by their PMI scores. [sent-113, score-0.432]
30 1Reisinger and Pas ¸ca (2009) considered the related problem of finding the most appropriate class for each attribute; they take an existing ontology of concepts (WordNet) as a class hierarchy and use a Bayesian approach to decide “the correct level of abstraction for each attribute. [sent-114, score-0.206]
31 ” Filtering Attributes We experimented with two different methods to select a final set of distinguishing attributes for each class: (1) we used a threshold to select the top-ranked attributes for each class, and (2) we manually filtered the attributes. [sent-115, score-0.971]
32 For the gender classification task, we manually filtered the entire set of attributes to select around 1000 attributes that were judged to be discriminative (two thirds of which are female). [sent-116, score-1.138]
33 We make these filter attributes available online as an attachment to this article, available through the ACL Anthology. [sent-119, score-0.474]
34 In our gold-standard gender data (Section 5), however, every user has a homepage [by dataset construction]; we might therefore incorrectly classify every user as Male. [sent-122, score-0.527]
35 ” Since our approach requires manual in- volvement in the filtering of the attribute list, one might argue that one should simply manually enumerate the most relevant attributes directly. [sent-128, score-0.602]
36 Psychologists therefore such lists by pooling the responses generate across many participants: future work may compare our “automatically generate, manually prune” approach to soliciting attributes via crowdsourcing. [sent-131, score-0.432]
37 2 Table 2 gives examples of our extracted at- 2One can also view the work of manually filtering attributes as a kind of “feature labeling. [sent-132, score-0.468]
38 Our attributes also go beyond the traditional meronyms that were the target of earlier work. [sent-146, score-0.432]
39 As we discuss further in Related Work (Section 7), previous researchers have worried about a proper definition of parts or attributes and relied on human judgments for evaluation (Berland and Charniak, 1999; Girju et al. [sent-147, score-0.432]
40 our results suggest that trying to identify attributes is beneficial. [sent-153, score-0.432]
41 ” 4 Applying Class Attributes To classify users using the extracted attributes, we look for cases where users refer to such attributes in their first-person writings. [sent-154, score-0.63]
42 3 We found that users most often reveal their attributes in the possessive construction, “my X” where X is an attribute, quality or event that they possess (in a linguistic sense). [sent-156, score-0.531]
43 4 We therefore assign a user to a demographic category as follows: We first part-of-speech tag our data using CRFTagger (Phan, 2006) and then look for “my X” patterns where X is a sequence of tokens terminating in a noun, analogous to our 3http://trec. [sent-159, score-0.277]
44 5 attribute-extraction pattern (Section When a user uses such a “my X” construction, we match the filler X against our attribute lists for each class. [sent-167, score-0.344]
45 If the filler is on a list, we call it a selfdistinguishing attribute of a user. [sent-168, score-0.244]
46 Our first technique provides a simple way to use our identified self-distinguishing attributes in conjunction with a classifier trained on gold-standard data. [sent-170, score-0.478]
47 If the user has any self- distinguishing attributes, we assign the user to the corresponding class; otherwise, we trust the output of the classifier. [sent-171, score-0.381]
48 (2) Bootstrapped: Automatic Labeling of Training Examples Even without gold standard training data, we can use our self-distinguishing attributes to automatically bootstrap annotations. [sent-172, score-0.488]
49 We collect a large pool ofunlabeled users and their tweets, and we apply the ARules described above to label those users that have self-distinguishing attributes. [sent-173, score-0.198]
50 Once an example is auto-annotated, we delete the self-distinguishing attributes from the user’s content. [sent-174, score-0.432]
51 We use the following simple but 5While we used an “off the shelf” POS tagger in this work, we note that taggers optimized specifically for social media are now available and would likely have resulted in higher tagging accuracy (e. [sent-179, score-0.143]
52 6Note that while our target gender task presents mutuallyexclusive output classes, we can still train classifiers for other categories without clear opposites (e. [sent-183, score-0.258]
53 for labeling users as Parents or Doctors) by using the 1-class classification paradigm (Koppel and Schler, 2004). [sent-185, score-0.153]
54 5 Twitter Gender Prediction To test the use of self-distinguishing attributes in user classification, we apply our methods to the task of gender classification on Twitter. [sent-190, score-0.843]
55 Since “many of these [sites] have wellstructured profile pages [where users] must se- lect gender and other attributes from dropdown menus,” they were able to link these attributes to the Twitter users. [sent-199, score-1.084]
56 We filter non-English tweets from this corpus using the LID system of Bergsma et al. [sent-201, score-0.15]
57 We then filter users with <40 tweets and randomly divide the remaining users into 2282 training, 1140 development, and 1141 test examples. [sent-203, score-0.348]
58 For the Usr features, we add special beginning and ending characters to the username, and then create features for all character n-grams of length twoto-four in the modified username string. [sent-214, score-0.153]
59 9 billion tweets collected from 07/201 1 to 11/2012, and 80 million tweets collected from the followers of 10-thousand location and languagespecific Twitter feeds. [sent-220, score-0.216]
60 We filter this corpus as above, except we do not put any restrictions on the number of tweets needed per user. [sent-221, score-0.15]
61 We also filter any users that overlap with our gold standard data. [sent-222, score-0.197]
62 The decisions of our bootstrapping process reflect the true gender distribution; the auto-annotated data is 60. [sent-224, score-0.255]
63 Figure 1 shows that a wide range of self-distinguishing attributes are used in the auto-annotation process. [sent-227, score-0.432]
64 This is important because if only a few attributes are used (e. [sent-228, score-0.432]
65 714 Figure 1: Frequency with which attributes are used to auto-annotate examples in the bootstrapping proach. [sent-234, score-0.503]
66 The plot identifies some attributes and their corresponding ap- class (labeled via gender symbol). [sent-235, score-0.755]
67 2 Table 3: Classification accuracy (%) on gold standard test data for user gender prediction on Twitter 6 Results Our main classification results are presented in Table 3. [sent-243, score-0.467]
68 It is also possible to use the ARules as a stand-alone system rather than as a post-process, however the coverage is low: we find a distinguishing attribute in 18. [sent-255, score-0.277]
69 3% of the 695 Female instances in the test data, and make the cor7Note that it is possible to achieve even higher performance on gender classification in social media if you have further information about a user, such as their full first and last name (Burger et al. [sent-256, score-0.456]
70 This is important because having thousands of gold standard annotations for every possible user characterization task, in every domain and social media platform, is not realistic. [sent-265, score-0.463]
71 Combining the bootstrapped classifier with the gold standard annotations in the BootStacked model results in further gains in performance. [sent-266, score-0.322]
72 8 These results provide strong validation for both the inherent utility ofclass-attributes knowledge in user characterization and the effectiveness of our specific strategies for exploiting such knowledge. [sent-267, score-0.264]
73 Female users using betcha and xox in their writing). [sent-280, score-0.173]
74 features capture reasonable associations between gender classes and particular names (such as mike, tony, omar, etc. [sent-284, score-0.331]
75 ) and also between gender classes and common nouns (such as guy, dad, sir, etc. [sent-285, score-0.257]
76 Many recent papers have analyzed the language of social media users, along dimensions such as ethnicity (Eisenstein et al. [sent-298, score-0.143]
77 , 2010; Pennacchiotti and Popescu, 2011) and gender (Rao et al. [sent-302, score-0.22]
78 ” Berland and Charniak (1999) did have success using Hearststyle patterns for part-whole detection, which they attribute to their “very large corpus and the use of more refined statistical measures for ranking the output. [sent-309, score-0.228]
79 Indeed, Berland and Charniak (1999) attempted to filter out attributes that were regarded as qualities (like driveability) rather than parts (like steering wheels) by removing words ending with the suffixes -ness, -ing, and -ity. [sent-314, score-0.509]
80 In our work, such qualities are not filtered and are ultimately valuable in classification; for example, the attributes peak fertility and loveliness are highly 716 associated with females. [sent-315, score-0.467]
81 As subsequent research became more focused on applications, looser definitions of class attributes were adopted. [sent-316, score-0.535]
82 Almuhareb and Poesio (2004) automatically mined class attributes that include parts, qualities, and those with an “agentive” or “telic” role with the class. [sent-317, score-0.535]
83 Their extended set of attributes was shown to enable an improved representation of nouns for the purpose of clustering these nouns into semantic concepts. [sent-318, score-0.432]
84 (2005) define attributes as properties that can serve as focus words in questions about a target class; e. [sent-320, score-0.432]
85 director is an attribute of a movie since one might ask, “Who is the director of this movie? [sent-322, score-0.17]
86 ” Another line of research has been motivated by the observation that much of Internet search consists of people looking for values of various class attributes (Bellare et al. [sent-323, score-0.535]
87 By knowing the attributes of different classes, search engines can better recognize that queries such as “altitude guadalajara” or “population guadalajara” are seeking values for a particular city’s “altitude” and “population” attributes (Pa ¸sca and Van Durme, 2007). [sent-326, score-0.864]
88 (2008) compared instance-based and class-based patterns for broad-definition attribute extraction, and found both to be effective. [sent-328, score-0.228]
89 Experts can manually specify the attributes of entities, as in the WordNet project (Miller et al. [sent-330, score-0.432]
90 Others have automatically extracted attribute relations from dictionary definitions (Richardson et al. [sent-332, score-0.17]
91 ” While we are the first to exploit commonsense knowledge in user characterization, common sense has been applied to a range of other problems in natural language processing. [sent-340, score-0.187]
92 8 Conclusion We have proposed, developed and successfully evaluated a novel approach to user characterization based on exploiting knowledge of user class attributes. [sent-350, score-0.504]
93 The knowledge is obtained using a new algorithm that discovers distinguishing attributes of particular classes. [sent-351, score-0.539]
94 Our approach to discovering distinguishing attributes represents a significant new direction for research in class-attribute extraction, and provides a valuable bridge between the fields of user characterization and lexical knowledge extraction. [sent-352, score-0.803]
95 All techniques lead to significant improve717 ments over state-of-the-art supervised systems on the task of Twitter gender classification. [sent-354, score-0.26]
96 Acquisition of instance attributes via labeled and related instances. [sent-359, score-0.432]
97 Broadly improving user classification via communication-based name and location clustering on twitter. [sent-383, score-0.191]
98 What you seek is what you get: extraction of class attributes from query logs. [sent-552, score-0.535]
99 Weakly-supervised acquisition of open-domain classes and class attributes from web documents and query logs. [sent-557, score-0.572]
100 Hierarchical bayesian models for latent attribute detection in social media. [sent-583, score-0.245]
wordName wordTfidf (topN-words)
[('attributes', 0.432), ('gender', 0.22), ('bootstrapped', 0.22), ('durme', 0.21), ('burger', 0.183), ('female', 0.179), ('attribute', 0.17), ('usr', 0.149), ('berland', 0.149), ('male', 0.139), ('user', 0.137), ('arules', 0.131), ('characterization', 0.127), ('username', 0.116), ('twitter', 0.116), ('tweets', 0.108), ('distinguishing', 0.107), ('bow', 0.106), ('rao', 0.106), ('class', 0.103), ('van', 0.101), ('koppel', 0.101), ('users', 0.099), ('pas', 0.096), ('schler', 0.092), ('bergsma', 0.085), ('pennacchiotti', 0.085), ('almuhareb', 0.083), ('demographic', 0.082), ('social', 0.075), ('bootstacked', 0.075), ('media', 0.068), ('fink', 0.066), ('charniak', 0.066), ('alfonseca', 0.061), ('patterns', 0.058), ('males', 0.058), ('boyhood', 0.056), ('maiden', 0.056), ('schubert', 0.056), ('tokunaga', 0.056), ('shane', 0.056), ('gold', 0.056), ('girju', 0.055), ('classification', 0.054), ('popescu', 0.052), ('commonsense', 0.05), ('car', 0.047), ('moshe', 0.047), ('classifier', 0.046), ('icwsm', 0.046), ('sociolinguistic', 0.046), ('benjamin', 0.045), ('wife', 0.045), ('marius', 0.045), ('sca', 0.043), ('filter', 0.042), ('richardson', 0.041), ('supervised', 0.04), ('ca', 0.04), ('connor', 0.039), ('instances', 0.039), ('tweet', 0.039), ('yarowsky', 0.039), ('poesio', 0.039), ('classifiers', 0.038), ('altitude', 0.037), ('autoannotated', 0.037), ('betcha', 0.037), ('dowry', 0.037), ('fatherhood', 0.037), ('guadalajara', 0.037), ('handbag', 0.037), ('lenhart', 0.037), ('selfdistinguishing', 0.037), ('soboroff', 0.037), ('waitress', 0.037), ('xox', 0.037), ('associations', 0.037), ('features', 0.037), ('mukherjee', 0.037), ('guy', 0.037), ('filler', 0.037), ('fillers', 0.037), ('zaidan', 0.037), ('classes', 0.037), ('examples', 0.036), ('qualities', 0.035), ('pmi', 0.035), ('bootstrapping', 0.035), ('omar', 0.034), ('dirac', 0.033), ('wives', 0.033), ('sir', 0.033), ('owoputi', 0.033), ('boxers', 0.033), ('homepage', 0.033), ('argamon', 0.033), ('dad', 0.033), ('garera', 0.033)]
simIndex simValue paperId paperTitle
same-paper 1 0.99999756 373 acl-2013-Using Conceptual Class Attributes to Characterize Social Media Users
Author: Shane Bergsma ; Benjamin Van Durme
Abstract: We describe a novel approach for automatically predicting the hidden demographic properties of social media users. Building on prior work in common-sense knowledge acquisition from third-person text, we first learn the distinguishing attributes of certain classes of people. For example, we learn that people in the Female class tend to have maiden names and engagement rings. We then show that this knowledge can be used in the analysis of first-person communication; knowledge of distinguishing attributes allows us to both classify users and to bootstrap new training examples. Our novel approach enables substantial improvements on the widelystudied task of user gender prediction, ob- taining a 20% relative error reduction over the current state-of-the-art.
2 0.25792423 249 acl-2013-Models of Semantic Representation with Visual Attributes
Author: Carina Silberer ; Vittorio Ferrari ; Mirella Lapata
Abstract: We consider the problem of grounding the meaning of words in the physical world and focus on the visual modality which we represent by visual attributes. We create a new large-scale taxonomy of visual attributes covering more than 500 concepts and their corresponding 688K images. We use this dataset to train attribute classifiers and integrate their predictions with text-based distributional models of word meaning. We show that these bimodal models give a better fit to human word association data compared to amodal models and word representations based on handcrafted norming data.
3 0.16615802 148 acl-2013-Exploring Sentiment in Social Media: Bootstrapping Subjectivity Clues from Multilingual Twitter Streams
Author: Svitlana Volkova ; Theresa Wilson ; David Yarowsky
Abstract: We study subjective language media and create Twitter-specific lexicons via bootstrapping sentiment-bearing terms from multilingual Twitter streams. Starting with a domain-independent, highprecision sentiment lexicon and a large pool of unlabeled data, we bootstrap Twitter-specific sentiment lexicons, using a small amount of labeled data to guide the process. Our experiments on English, Spanish and Russian show that the resulting lexicons are effective for sentiment classification for many underexplored languages in social media.
4 0.15678367 115 acl-2013-Detecting Event-Related Links and Sentiments from Social Media Texts
Author: Alexandra Balahur ; Hristo Tanev
Abstract: Nowadays, the importance of Social Media is constantly growing, as people often use such platforms to share mainstream media news and comment on the events that they relate to. As such, people no loger remain mere spectators to the events that happen in the world, but become part of them, commenting on their developments and the entities involved, sharing their opinions and distributing related content. This paper describes a system that links the main events detected from clusters of newspaper articles to tweets related to them, detects complementary information sources from the links they contain and subsequently applies sentiment analysis to classify them into positive, negative and neutral. In this manner, readers can follow the main events happening in the world, both from the perspective of mainstream as well as social media and the public’s perception on them. This system will be part of the EMM media monitoring framework working live and it will be demonstrated using Google Earth.
5 0.13189545 20 acl-2013-A Stacking-based Approach to Twitter User Geolocation Prediction
Author: Bo Han ; Paul Cook ; Timothy Baldwin
Abstract: We implement a city-level geolocation prediction system for Twitter users. The system infers a user’s location based on both tweet text and user-declared metadata using a stacking approach. We demonstrate that the stacking method substantially outperforms benchmark methods, achieving 49% accuracy on a benchmark dataset. We further evaluate our method on a recent crawl of Twitter data to investigate the impact of temporal factors on model generalisation. Our results suggest that user-declared location metadata is more sensitive to temporal change than the text of Twitter messages. We also describe two ways of accessing/demoing our system.
6 0.1251125 168 acl-2013-Generating Recommendation Dialogs by Extracting Information from User Reviews
8 0.10655279 240 acl-2013-Microblogs as Parallel Corpora
9 0.1003947 114 acl-2013-Detecting Chronic Critics Based on Sentiment Polarity and Userâ•Žs Behavior in Social Media
10 0.096989959 33 acl-2013-A user-centric model of voting intention from Social Media
11 0.092915818 233 acl-2013-Linking Tweets to News: A Framework to Enrich Short Text Data in Social Media
12 0.092721045 180 acl-2013-Handling Ambiguities of Bilingual Predicate-Argument Structures for Statistical Machine Translation
13 0.089914888 90 acl-2013-Conditional Random Fields for Responsive Surface Realisation using Global Features
14 0.088822037 45 acl-2013-An Empirical Study on Uncertainty Identification in Social Media Context
15 0.08665821 319 acl-2013-Sequential Summarization: A New Application for Timely Updated Twitter Trending Topics
16 0.084428437 169 acl-2013-Generating Synthetic Comparable Questions for News Articles
17 0.083356231 147 acl-2013-Exploiting Topic based Twitter Sentiment for Stock Prediction
18 0.082837015 356 acl-2013-Transfer Learning Based Cross-lingual Knowledge Extraction for Wikipedia
19 0.081218943 306 acl-2013-SPred: Large-scale Harvesting of Semantic Predicates
20 0.077491201 121 acl-2013-Discovering User Interactions in Ideological Discussions
topicId topicWeight
[(0, 0.208), (1, 0.158), (2, -0.015), (3, 0.006), (4, 0.044), (5, 0.002), (6, 0.088), (7, 0.023), (8, 0.138), (9, -0.038), (10, -0.15), (11, -0.045), (12, 0.043), (13, 0.023), (14, 0.007), (15, -0.034), (16, 0.021), (17, 0.062), (18, 0.029), (19, -0.024), (20, 0.007), (21, 0.063), (22, 0.043), (23, -0.049), (24, 0.12), (25, 0.005), (26, 0.061), (27, 0.004), (28, 0.016), (29, 0.022), (30, -0.059), (31, 0.064), (32, 0.053), (33, 0.092), (34, -0.008), (35, 0.062), (36, -0.003), (37, -0.048), (38, 0.007), (39, 0.001), (40, -0.027), (41, 0.106), (42, -0.0), (43, -0.006), (44, 0.029), (45, -0.045), (46, 0.006), (47, -0.014), (48, -0.004), (49, 0.006)]
simIndex simValue paperId paperTitle
same-paper 1 0.9350819 373 acl-2013-Using Conceptual Class Attributes to Characterize Social Media Users
Author: Shane Bergsma ; Benjamin Van Durme
Abstract: We describe a novel approach for automatically predicting the hidden demographic properties of social media users. Building on prior work in common-sense knowledge acquisition from third-person text, we first learn the distinguishing attributes of certain classes of people. For example, we learn that people in the Female class tend to have maiden names and engagement rings. We then show that this knowledge can be used in the analysis of first-person communication; knowledge of distinguishing attributes allows us to both classify users and to bootstrap new training examples. Our novel approach enables substantial improvements on the widelystudied task of user gender prediction, ob- taining a 20% relative error reduction over the current state-of-the-art.
2 0.7719487 20 acl-2013-A Stacking-based Approach to Twitter User Geolocation Prediction
Author: Bo Han ; Paul Cook ; Timothy Baldwin
Abstract: We implement a city-level geolocation prediction system for Twitter users. The system infers a user’s location based on both tweet text and user-declared metadata using a stacking approach. We demonstrate that the stacking method substantially outperforms benchmark methods, achieving 49% accuracy on a benchmark dataset. We further evaluate our method on a recent crawl of Twitter data to investigate the impact of temporal factors on model generalisation. Our results suggest that user-declared location metadata is more sensitive to temporal change than the text of Twitter messages. We also describe two ways of accessing/demoing our system.
3 0.73992538 45 acl-2013-An Empirical Study on Uncertainty Identification in Social Media Context
Author: Zhongyu Wei ; Junwen Chen ; Wei Gao ; Binyang Li ; Lanjun Zhou ; Yulan He ; Kam-Fai Wong
Abstract: Uncertainty text detection is important to many social-media-based applications since more and more users utilize social media platforms (e.g., Twitter, Facebook, etc.) as information source to produce or derive interpretations based on them. However, existing uncertainty cues are ineffective in social media context because of its specific characteristics. In this paper, we propose a variant of annotation scheme for uncertainty identification and construct the first uncertainty corpus based on tweets. We then conduct experiments on the generated tweets corpus to study the effectiveness of different types of features for uncertainty text identification.
4 0.73939395 33 acl-2013-A user-centric model of voting intention from Social Media
Author: Vasileios Lampos ; Daniel Preoţiuc-Pietro ; Trevor Cohn
Abstract: Social Media contain a multitude of user opinions which can be used to predict realworld phenomena in many domains including politics, finance and health. Most existing methods treat these problems as linear regression, learning to relate word frequencies and other simple features to a known response variable (e.g., voting intention polls or financial indicators). These techniques require very careful filtering of the input texts, as most Social Media posts are irrelevant to the task. In this paper, we present a novel approach which performs high quality filtering automatically, through modelling not just words but also users, framed as a bilinear model with a sparse regulariser. We also consider the problem of modelling groups of related output variables, using a structured multi-task regularisation method. Our experiments on voting intention prediction demonstrate strong performance over large-scale input from Twitter on two distinct case studies, outperforming competitive baselines.
Author: Simone Paolo Ponzetto ; Andrea Zielinski
Abstract: unkown-abstract
6 0.6568023 95 acl-2013-Crawling microblogging services to gather language-classified URLs. Workflow and case study
7 0.65478307 115 acl-2013-Detecting Event-Related Links and Sentiments from Social Media Texts
8 0.65257293 114 acl-2013-Detecting Chronic Critics Based on Sentiment Polarity and Userâ•Žs Behavior in Social Media
9 0.6373328 42 acl-2013-Aid is Out There: Looking for Help from Tweets during a Large Scale Disaster
10 0.61793768 301 acl-2013-Resolving Entity Morphs in Censored Data
11 0.58259821 30 acl-2013-A computational approach to politeness with application to social factors
12 0.58025253 356 acl-2013-Transfer Learning Based Cross-lingual Knowledge Extraction for Wikipedia
13 0.57005662 233 acl-2013-Linking Tweets to News: A Framework to Enrich Short Text Data in Social Media
14 0.56702834 141 acl-2013-Evaluating a City Exploration Dialogue System with Integrated Question-Answering and Pedestrian Navigation
15 0.54436129 249 acl-2013-Models of Semantic Representation with Visual Attributes
16 0.53968918 346 acl-2013-The Impact of Topic Bias on Quality Flaw Prediction in Wikipedia
17 0.53775978 298 acl-2013-Recognizing Rare Social Phenomena in Conversation: Empowerment Detection in Support Group Chatrooms
18 0.53465652 340 acl-2013-Text-Driven Toponym Resolution using Indirect Supervision
19 0.51854628 268 acl-2013-PATHS: A System for Accessing Cultural Heritage Collections
20 0.51669049 168 acl-2013-Generating Recommendation Dialogs by Extracting Information from User Reviews
topicId topicWeight
[(0, 0.052), (6, 0.032), (11, 0.053), (15, 0.016), (24, 0.065), (26, 0.082), (28, 0.013), (35, 0.068), (42, 0.031), (44, 0.215), (48, 0.044), (70, 0.081), (71, 0.014), (88, 0.045), (90, 0.02), (95, 0.062)]
simIndex simValue paperId paperTitle
same-paper 1 0.81367964 373 acl-2013-Using Conceptual Class Attributes to Characterize Social Media Users
Author: Shane Bergsma ; Benjamin Van Durme
Abstract: We describe a novel approach for automatically predicting the hidden demographic properties of social media users. Building on prior work in common-sense knowledge acquisition from third-person text, we first learn the distinguishing attributes of certain classes of people. For example, we learn that people in the Female class tend to have maiden names and engagement rings. We then show that this knowledge can be used in the analysis of first-person communication; knowledge of distinguishing attributes allows us to both classify users and to bootstrap new training examples. Our novel approach enables substantial improvements on the widelystudied task of user gender prediction, ob- taining a 20% relative error reduction over the current state-of-the-art.
2 0.81043637 280 acl-2013-Plurality, Negation, and Quantification:Towards Comprehensive Quantifier Scope Disambiguation
Author: Mehdi Manshadi ; Daniel Gildea ; James Allen
Abstract: Recent work on statistical quantifier scope disambiguation (QSD) has improved upon earlier work by scoping an arbitrary number and type of noun phrases. No corpusbased method, however, has yet addressed QSD when incorporating the implicit universal of plurals and/or operators such as negation. In this paper we report early, though promising, results for automatic QSD when handling both phenomena. We also present a general model for learning to build partial orders from a set of pairwise preferences. We give an n log n algorithm for finding a guaranteed approximation of the optimal solution, which works very well in practice. Finally, we significantly improve the performance of the pre- vious model using a rich set of automatically generated features.
3 0.70147401 134 acl-2013-Embedding Semantic Similarity in Tree Kernels for Domain Adaptation of Relation Extraction
Author: Barbara Plank ; Alessandro Moschitti
Abstract: Relation Extraction (RE) is the task of extracting semantic relationships between entities in text. Recent studies on relation extraction are mostly supervised. The clear drawback of supervised methods is the need of training data: labeled data is expensive to obtain, and there is often a mismatch between the training data and the data the system will be applied to. This is the problem of domain adaptation. In this paper, we propose to combine (i) term generalization approaches such as word clustering and latent semantic analysis (LSA) and (ii) structured kernels to improve the adaptability of relation extractors to new text genres/domains. The empirical evaluation on ACE 2005 domains shows that a suitable combination of syntax and lexical generalization is very promising for domain adaptation.
4 0.63987374 318 acl-2013-Sentiment Relevance
Author: Christian Scheible ; Hinrich Schutze
Abstract: A number of different notions, including subjectivity, have been proposed for distinguishing parts of documents that convey sentiment from those that do not. We propose a new concept, sentiment relevance, to make this distinction and argue that it better reflects the requirements of sentiment analysis systems. We demonstrate experimentally that sentiment relevance and subjectivity are related, but different. Since no large amount of labeled training data for our new notion of sentiment relevance is available, we investigate two semi-supervised methods for creating sentiment relevance classifiers: a distant supervision approach that leverages structured information about the domain of the reviews; and transfer learning on feature representations based on lexical taxonomies that enables knowledge transfer. We show that both methods learn sentiment relevance classifiers that perform well.
5 0.63696563 369 acl-2013-Unsupervised Consonant-Vowel Prediction over Hundreds of Languages
Author: Young-Bum Kim ; Benjamin Snyder
Abstract: In this paper, we present a solution to one aspect of the decipherment task: the prediction of consonants and vowels for an unknown language and alphabet. Adopting a classical Bayesian perspective, we performs posterior inference over hundreds of languages, leveraging knowledge of known languages and alphabets to uncover general linguistic patterns of typologically coherent language clusters. We achieve average accuracy in the unsupervised consonant/vowel prediction task of 99% across 503 languages. We further show that our methodology can be used to predict more fine-grained phonetic distinctions. On a three-way classification task between vowels, nasals, and nonnasal consonants, our model yields unsu- pervised accuracy of 89% across the same set of languages.
6 0.63616472 144 acl-2013-Explicit and Implicit Syntactic Features for Text Classification
7 0.63087428 2 acl-2013-A Bayesian Model for Joint Unsupervised Induction of Sentiment, Aspect and Discourse Representations
8 0.62763536 169 acl-2013-Generating Synthetic Comparable Questions for News Articles
9 0.62142807 7 acl-2013-A Lattice-based Framework for Joint Chinese Word Segmentation, POS Tagging and Parsing
10 0.61847353 131 acl-2013-Dual Training and Dual Prediction for Polarity Classification
11 0.6182369 272 acl-2013-Paraphrase-Driven Learning for Open Question Answering
12 0.6166755 233 acl-2013-Linking Tweets to News: A Framework to Enrich Short Text Data in Social Media
13 0.61574155 80 acl-2013-Chinese Parsing Exploiting Characters
14 0.61454529 224 acl-2013-Learning to Extract International Relations from Political Context
15 0.61429566 249 acl-2013-Models of Semantic Representation with Visual Attributes
16 0.61384273 82 acl-2013-Co-regularizing character-based and word-based models for semi-supervised Chinese word segmentation
17 0.61266088 81 acl-2013-Co-Regression for Cross-Language Review Rating Prediction
18 0.61130774 346 acl-2013-The Impact of Topic Bias on Quality Flaw Prediction in Wikipedia
19 0.61129272 254 acl-2013-Multimodal DBN for Predicting High-Quality Answers in cQA portals
20 0.61117917 274 acl-2013-Parsing Graphs with Hyperedge Replacement Grammars