Unlabeled Tweets Research Articles

ACOVMD: Automatic COVID‐19 misinformation detection in Twitter using self‐trained semi‐supervised hybrid deep learning model

AbstractDuring the COVID‐19 pandemic, online social networks are extensively utilized, more than ever before by 8.4%, resulting in the propagation of false information related to COVID‐19. Despite the existence of many fake news detection models; annotation inconsistency, memory consumption, accurate and self‐trained efficient algorithms for detecting the emerging COVID‐19 misinformation tweets are still challenging. Hence, the main aim of this work is to come up with a self‐trained semi‐supervised model that accurately and automatically detects the reliability of emerging COVID‐19 tweets without delay. In this work, COVID‐19 tweet dataset is created in English Language from the period January 2020 to January 2022 as a ground truth database. Then self‐trained semi‐supervised hybrid deep learning model is proposed to train both supervised and unsupervised components simultaneously using the created dataset. The proposed model is self‐trained repeatedly and the model gets updated to predict the reliability of upcoming COVID‐19 tweets that differ from training tweets. We performed experiments multiple times by limiting the percentage amount of labelled tweets shown to the model, namely 80%, 50%, 40%, 30%, 20% and 10% labelled tweets, respectively. Experimental results show that the proposed model achieves 80.92% accuracy and 98.15% accuracy in the 10% and 80% label‐seen experiments, respectively. This shows a clear rising trend in the performance curve. Therefore, this technique will be useful for effectively classifying voluminous amounts of emerging tweets generated as part of the COVID‐19 infodemic. The proposed model may efficiently use a huge amount of unlabelled tweets and enhance the model's generalization performance.

Q8VaxStance: Dataset Labeling System for Stance Detection towards Vaccines in Kuwaiti Dialect

The Kuwaiti dialect is a particular dialect of Arabic spoken in Kuwait; it differs significantly from standard Arabic and the dialects of neighboring countries in the same region. Few research papers with a focus on the Kuwaiti dialect have been published in the field of NLP. In this study, we created Kuwaiti dialect language resources using Q8VaxStance, a vaccine stance labeling system for a large dataset of tweets. This dataset fills this gap and provides a valuable resource for researchers studying vaccine hesitancy in Kuwait. Furthermore, it contributes to the Arabic natural language processing field by providing a dataset for developing and evaluating machine learning models for stance detection in the Kuwaiti dialect. The proposed vaccine stance labeling system combines the benefits of weak supervised learning and zero-shot learning; for this purpose, we implemented 52 experiments on 42,815 unlabeled tweets extracted between December 2020 and July 2022. The results of the experiments show that using keyword detection in conjunction with zero-shot model labeling functions is significantly better than using only keyword detection labeling functions or just zero-shot model labeling functions. Furthermore, for the total number of generated labels, the difference between using the Arabic language in both the labels and prompt or a mix of Arabic labels and an English prompt is statistically significant, indicating that it generates more labels than when using English in both the labels and prompt. The best accuracy achieved in our experiments in terms of the Macro-F1 values was found when using keyword and hashtag detection labeling functions in conjunction with zero-shot model labeling functions, specifically in experiments KHZSLF-EE4 and KHZSLF-EA1, with values of 0.83 and 0.83, respectively. Experiment KHZSLF-EE4 was able to label 42,270 tweets, while experiment KHZSLF-EA1 was able to label 42,764 tweets. Finally, the average value of annotation agreement between the generated labels and human labels ranges between 0.61 and 0.64, which is considered a good level of agreement.

Automatically Identifying Twitter Users for Interventions to Support Dementia Family Caregivers: Annotated Data Set and Benchmark Classification Models.

BackgroundMore than 6 million people in the United States have Alzheimer disease and related dementias, receiving help from more than 11 million family or other informal caregivers. A range of traditional interventions has been developed to support family caregivers; however, most of them have not been implemented in practice and remain largely inaccessible. While recent studies have shown that family caregivers of people with dementia use Twitter to discuss their experiences, methods have not been developed to enable the use of Twitter for interventions.ObjectiveThe objective of this study is to develop an annotated data set and benchmark classification models for automatically identifying a cohort of Twitter users who have a family member with dementia.MethodsBetween May 4 and May 20, 2021, we collected 10,733 tweets, posted by 8846 users, that mention a dementia-related keyword, a linguistic marker that potentially indicates a diagnosis, and a select familial relationship. Three annotators annotated 1 random tweet per user to distinguish those that indicate having a family member with dementia from those that do not. Interannotator agreement was 0.82 (Fleiss kappa). We used the annotated tweets to train and evaluate support vector machine and deep neural network classifiers. To assess the scalability of our approach, we then deployed automatic classification on unlabeled tweets that were continuously collected between May 4, 2021, and March 9, 2022.ResultsA deep neural network classifier based on a BERT (bidirectional encoder representations from transformers) model pretrained on tweets achieved the highest F1-score of 0.962 (precision=0.946 and recall=0.979) for the class of tweets indicating that the user has a family member with dementia. The classifier detected 128,838 tweets that indicate having a family member with dementia, posted by 74,290 users between May 4, 2021, and March 9, 2022—that is, approximately 7500 users per month.ConclusionsOur annotated data set can be used to automatically identify Twitter users who have a family member with dementia, enabling the use of Twitter on a large scale to not only explore family caregivers’ experiences but also directly target interventions at these users.

Weakly labeled data augmentation for social media named entity recognition

Named entity recognition is a task that extracts entities corresponding to predefined categories. Although NER is important in processing user-generated texts such as those obtained from social media, it remains challenging because such texts tend to contain numerous unseen words or abbreviations. To address this issue, we propose two methods for weakly labeled data generation that can extract named entities from social media texts more effectively: alias augmentation and typo augmentation. Using these methods, weakly labeled data are generated through the automatic annotation of unlabeled Wikipedia texts and Tweets and then trained through transfer learning. Our experimental results suggest that the proposed approach improves NER performance, with our best F1-score of 51.43% representing the highest score ever reported.

A Novel Annotation Scheme to Generate Hate Speech Corpus through Crowdsourcing and Active Learning

The number of user-generated posts is growing exponentially with social media usage growth. Promoting violence against or having the primary purpose of inciting hatred against individuals or groups based on specific attributes via social media posts is daunting. As the posts are published in multiple languages with different forms of multimedia, social media finds it challenging to moderate before reaching the audience and assessing the posts as hate speech becomes sophisticated due to subjectivity. Social media platforms lack contextual and linguistic expertise and social and cultural insights to identify hate speech accurately. Research is being carried out to detect hate speech on social media content in English using machine learning algorithms, etc., using different crowdsourcing platforms. However, these platforms' workers are unavailable from countries such as Sri Lanka. The lack of a workforce with the necessary skill set and annotation schemes symbolizes further research essentiality in low-resource language annotation. This research proposes a suitable crowdsourcing approach to label and annotates social media content to generate corpora with words and phrases to identify hate speech using machine learning algorithms in Sri Lanka. This paper summarizes the annotated Facebook posts, comments, and replies to comments from public Sri Lankan Facebook user profiles, pages and groups of 52,646 instances, unlabeled tweets based on 996 Twitter search keywords of 45,000 instances of YouTube Videos of 45,000 instances using the proposed annotation scheme. 9%, 21% and 14% of Facebook, Twitter and YouTube posts were identified as containing hate content. In addition, the posts were categorized as offensive and non-offensive, and hate targets and corpus associated with hate targets focusing on an individual or group were identified and presented in this paper. The proposed annotation scheme could be extended to other low-resource languages to identify the hate speech corpora. With the use of a well-implemented crowdsourcing platform with the proposed novel annotation scheme, it will be possible to find more subtle patterns with human judgment and filtering and take preventive measures to create a better cyberspace.

Extracting Diverse Sentiment Expressions with Target-Dependent Polarity from Twitter

The problem of automatic extraction of sentiment expressions from informal text, as in microblogs such as tweets is a recent area of investigation. Compared to formal text, such as in product reviews or news articles, one of the key challenges lies in the wide diversity and informal nature of sentiment expressions that cannot be trivially enumerated or captured using predefined lexical patterns. In this work, we present an optimization-based approach to automatically extract sentiment expressions for a given target (e.g., movie, or person) from a corpus of unlabeled tweets. Specifically, we make three contributions: (i) we recognize a diverse and richer set of sentiment-bearing expressions in tweets, including formal and slang words/phrases, not limited to pre-specified syntactic patterns; (ii) instead of associating sentiment with an entire tweet, we assess the target-dependent polarity of each sentiment expression. The polarity of sentiment expression is determined by the nature of its target; (iii) we provide a novel formulation of assigning polarity to a sentiment expression as a constrained optimization problem over the tweet corpus. Experiments conducted on two domains, tweets mentioning movie and person entities, show that our approach improves accuracy in comparison with several baseline methods, and that the improvement becomes more prominent with increasing corpus sizes.

Identifying Covid-19 misinformation tweets and learning their spatio-temporal topic dynamics using Nonnegative Coupled Matrix Tensor Factorization.

Social media platforms like Twitter have become an easy portal for billions of people to connect and exchange their thoughts. Unfortunately, people commonly use these platforms to share misinformation which can influence other people adversely. The spread of misinformation is unavoidable in an extraordinary situation like Covid-19, and the consequences can be dreadful. This paper proposes a two-step ranking-based misinformation detection (RMiD) technique. Firstly, a novel ranking-based approach leveraging the scalable information retrieval infrastructure is applied to detect misinformation from a huge collection of unlabelled tweets based on a related but very small labelled misinformation data set. Secondly, the identified misinformation tweets are represented as a coupled matrix tensor model and Nonnegative Coupled Matrix Tensor Factorization is applied to learn their spatio-temporal topic dynamics. The experimental analysis shows that RMiD is capable of detecting misinformation with better coverage and less noise in comparison with existing techniques. Moreover, the coupled matrix tensor representation has improved the quality of topics discovered from unlabelled data up to 4% by leveraging the semantic similarity of terms in labelled data.Supplementary InformationThe online version supplementary material available at 10.1007/s13278-021-00767-7.

COVID-19 outbreak: An ensemble pre-trained deep learning model for detecting informative tweets

On 11 March 2020, the (WHO) World Health Organization declared COVID-19 (CoronaVirus Disease 2019) as a pandemic. A further crisis has manifested mass fear and panic, driven by lack of information, or sometimes outright misinformation, alongside the coronavirus pandemic. Twitter is one of the prominent and trusted social media in this current outbreak. Over time, boundless COVID-19 headlines and vast awareness have been spreading, with tweets, updates, videos, and explosive posts. Few studies have been performed on the pandemic to detect and interrelate various disease types, including current coronavirus. However, it is pretty tricky to discriminate and detect a specific category. This work is motivated by the need to inform society about limiting irrelevant information and avoiding spreading negative emotions. In this context, the current work focuses on informative tweet detection in the pandemic to provide relevant information to the government, medical organizations, victims services, etc. This paper used a Majority Voting technique-based Ensemble Deep Learning (MVEDL) model. This MVEDL model is used to identify COVID-19 related (INFORMATIVE) tweets. The state-of-art deep learning models RoBERTa, BERTweet, and CT-BERT are used for best performance with the MVEDL model. The “COVID-19 English labeled tweets” dataset is used for training and testing the MVEDL model. The MVEDL model has shown 91.75 percent accuracy, 91.14 percent F1-score and outperforms the traditional machine learning and deep learning models. We also investigate how to use the MVEDL model for sentiment analysis on 226668 unlabeled COVID-19 tweets and their informative tweets. The application section discussed a comprehensive analysis of both actual and informative tweets. According to our knowledge, this is the first work on COVID-19 sentiment analysis using a deep learning ensemble model.

An enhanced cosine-based visual technique for the robust tweets data clustering

PurposeThe purpose of the paper is to study multiple viewpoints which are required to access the more informative similarity features among the tweets documents, which is useful for achieving the robust tweets data clustering results.Design/methodology/approachLet “N” be the number of tweets documents for the topics extraction. Unwanted texts, punctuations and other symbols are removed, tokenization and stemming operations are performed in the initial tweets pre-processing step. Bag-of-features are determined for the tweets; later tweets are modelled with the obtained bag-of-features during the process of topics extraction. Approximation of topics features are extracted for every tweet document. These set of topics features of N documents are treated as multi-viewpoints. The key idea of the proposed work is to use multi-viewpoints in the similarity features computation. The following figure illustrates multi-viewpoints based cosine similarity computation of the five tweets documents (here N = 5) and corresponding documents are defined in projected space with five viewpoints, say, v1,v2, v3, v4, and v5. For example, similarity features between two documents (viewpoints v1, and v2) are computed concerning the other three multi-viewpoints (v3, v4, and v5), unlike a single viewpoint in traditional cosine metric.FindingsHealthcare problems with tweets data. Topic models play a crucial role in the classification of health-related tweets with finding topics (or health clusters) instead of finding term frequency and inverse document frequency (TF–IDF) for unlabelled tweets.Originality/valueTopic models play a crucial role in the classification of health-related tweets with finding topics (or health clusters) instead of finding TF-IDF for unlabelled tweets.

Toward Using Twitter for Tracking COVID-19: A Natural Language Processing Pipeline and Exploratory Data Set.

BackgroundIn the United States, the rapidly evolving COVID-19 outbreak, the shortage of available testing, and the delay of test results present challenges for actively monitoring its spread based on testing alone.ObjectiveThe objective of this study was to develop, evaluate, and deploy an automatic natural language processing pipeline to collect user-generated Twitter data as a complementary resource for identifying potential cases of COVID-19 in the United States that are not based on testing and, thus, may not have been reported to the Centers for Disease Control and Prevention.MethodsBeginning January 23, 2020, we collected English tweets from the Twitter Streaming application programming interface that mention keywords related to COVID-19. We applied handwritten regular expressions to identify tweets indicating that the user potentially has been exposed to COVID-19. We automatically filtered out “reported speech” (eg, quotations, news headlines) from the tweets that matched the regular expressions, and two annotators annotated a random sample of 8976 tweets that are geo-tagged or have profile location metadata, distinguishing tweets that self-report potential cases of COVID-19 from those that do not. We used the annotated tweets to train and evaluate deep neural network classifiers based on bidirectional encoder representations from transformers (BERT). Finally, we deployed the automatic pipeline on more than 85 million unlabeled tweets that were continuously collected between March 1 and August 21, 2020.ResultsInterannotator agreement, based on dual annotations for 3644 (41%) of the 8976 tweets, was 0.77 (Cohen κ). A deep neural network classifier, based on a BERT model that was pretrained on tweets related to COVID-19, achieved an F1-score of 0.76 (precision=0.76, recall=0.76) for detecting tweets that self-report potential cases of COVID-19. Upon deploying our automatic pipeline, we identified 13,714 tweets that self-report potential cases of COVID-19 and have US state–level geolocations.ConclusionsWe have made the 13,714 tweets identified in this study, along with each tweet’s time stamp and US state–level geolocation, publicly available to download. This data set presents the opportunity for future work to assess the utility of Twitter data as a complementary resource for tracking the spread of COVID-19.

Improved semi-supervised learning technique for automatic detection of South African abusive language on Twitter

Semi-supervised learning is a potential solution for improving training data in low-resourced abusive language detection contexts such as South African abusive language detection on Twitter. However, the existing semi-supervised learning methods have been skewed towards small amounts of labelled data, with small feature space. This paper, therefore, presents a semi-supervised learning technique that improves the distribution of training data by assigning labels to unlabelled data based on the majority voting over different feature sets of labelled and unlabelled data clusters. The technique is applied to South African English corpora consisting of labelled and unlabelled abusive tweets. The proposed technique is compared with state-of-the-art self-learning and active learning techniques based on syntactic and semantic features. The performance of these techniques with Logistic Regression, Support Vector Machine and Neural Networks are evaluated. The proposed technique, with accuracy and F1-score of 0.97 and 0.95, respectively, outperforms existing semi-supervised learning techniques. The learning curves show that the training data was used more efficiently by the proposed technique compared to existing techniques. Overall, n-gram syntactic features with a Logistic Regression classifier records the highest performance. The paper concludes that the proposed semi-supervised learning technique effectively detected implicit and explicit South African abusive language on Twitter.

Towards countering hate speech against journalists on social media

The damaging effects of hate speech on social media are evident during the last few years, and several organizations, researchers and social media platforms tried to harness them in various ways. Despite these efforts, social media users are still affected by hate speech. The problem is even more apparent to social groups that promote public discourse, such as journalists. In this work, we focus on countering hate speech that is targeted to journalistic social media accounts. To accomplish this, a group of journalists assembled a definition of hate speech, taking into account the journalistic point of view and the types of hate speech that are usually targeted against journalists. We then compile a large pool of tweets referring to journalism-related accounts in multiple languages. In order to annotate the pool of unlabeled tweets according to the definition, we follow a concise annotation strategy that involves active learning annotation stages. The outcome of this paper is a novel, publicly available collection of Twitter datasets in five different languages. Additionally, we experiment with state-of-the-art deep learning architectures for hate speech detection and use our annotated datasets to train and evaluate them. Finally, we propose an ensemble detection model that outperforms all individual models.

An Active Learning Based Emoji Prediction Method in Turkish

Emoji usage has become a standard in social media platforms since it can condense feelings beyond short textual information. Recent advances in machine learning enable to write short messages with automatically detected emojis. However, the prediction of emojis for the given short message can be complicated, inasmuch as users can interpret different meanings beyond the intent of their designers. Therefore, an automatic extraction strategy of training samples cannot be convenient from the large volumes of unlabeled tweets. In this paper, we present an active learning method to evaluate the emoji prediction of a tweet with a limited number of labeled Turkish emoji dataset. To simulate a human-machine collaborative learning method, we train an initial classifier with this dataset and then we update the classifier by filtering related samples out from the large pool of unlabeled data. In the evaluation, we split 25\% randomly selected tweets combined with only one emoji from the generated dataset as a test case. Our active learning method has achieved 0.901 F1 score and outperforms other baseline supervised learning methods.

Personalized spoiler detection in tweets by using support vector machine

At present, numerous people watch prerecorded TV programs as daily leisure. Concerning soap operas or sports, the viewers may not want to be informed about the results before watching the programs; however, they may check tweets on devices, such as smartphones, some of which can accidentally include contents referring to spoilers. To avoid reading such contents, several approaches were proposed to detect spoilers in texts (both long and short ones), including tweets. In the study by Jeon et al. focused on detecting spoilers in tweets, only one person attached labels to tweets, and the labeled tweets were used to train detectors. The trained detector was tuned for one person and, therefore, could be unsuitable for others. A tweet published in the middle of a baseball game can be considered as a spoiler by some people and is not by others; therefore, a personalized detection method is preferred. However, to the best of our knowledge, none of the related studies has considered such a personalized approach. To address this problem, we propose a semi-supervised approach to detect spoilers in tweets using a Support Vector Machine (SVM) in which each user attaches labels to tweets. After that, SVM executes the same procedure for other unlabeled tweets through bootstrapping. To verify the suitability of the proposed approach to personalize detectors, we conducted an experiment in which two participants were asked to attach labels to tweets. The experimental results indicate that this approach is ef??icient for personalized detection based on the Mann-Whitney U test.

Towards scaling Twitter for digital epidemiology of birth defects

Social media has recently been used to identify and study a small cohort of Twitter users whose pregnancies with birth defect outcomes—the leading cause of infant mortality—could be observed via their publicly available tweets. In this study, we exploit social media on a larger scale by developing natural language processing (NLP) methods to automatically detect, among thousands of users, a cohort of mothers reporting that their child has a birth defect. We used 22,999 annotated tweets to train and evaluate supervised machine learning algorithms—feature-engineered and deep learning-based classifiers—that automatically distinguish tweets referring to the user’s pregnancy outcome from tweets that merely mention birth defects. Because 90% of the tweets merely mention birth defects, we experimented with under-sampling and over-sampling approaches to address this class imbalance. An SVM classifier achieved the best performance for the two positive classes: an F1-score of 0.65 for the “defect” class and 0.51 for the “possible defect” class. We deployed the classifier on 20,457 unlabeled tweets that mention birth defects, which helped identify 542 additional users for potential inclusion in our cohort. Contributions of this study include (1) NLP methods for automatically detecting tweets by users reporting their birth defect outcomes, (2) findings that an SVM classifier can outperform a deep neural network-based classifier for highly imbalanced social media data, (3) evidence that automatic classification can be used to identify additional users for potential inclusion in our cohort, and (4) a publicly available corpus for training and evaluating supervised machine learning algorithms.

Prediction of Personal Experience Tweets of Medication Use via Contextual Word Representations.

Continuous monitoring the safe use of medication is an important task in pharmacovigilance. The first-hand experiences of medication effects come from the consumers of the pharmaceuticals. Social media have been considered as a possible alternative data source for gathering consumer-generated information of their experience with medications. Identifying personal experience in social media data is a challenging task in natural language processing. In this study, we investigated a method of predicating personal experience tweets using Google's Bidirectional Encoder Representations from Transformers (BERT) and neural networks, in which BERT models contextually represented the tweet text. Both pre-trained BERT models and our BERT model trained with 3.2 million unlabeled tweets were examined. Our results show that our trained BERT model performs better than Google's pre-trained models (p <; 0.01). This suggests that domain-specific data may contribute to the BERT model yielding better classification performance in predicting personal experience tweets of medication use.

Unsupervised collective-based framework for dynamic retraining of supervised real-time spam tweets detection model

Twitter is one of the most popular social platforms. It has changed the way of communication and information dissemination through its real-time messaging mechanism. Recently, it has been used by researchers and industries as a new source of data for various intelligent systems, such as tweet sentiment analysis and recommendation systems, which require high data quality. However, due to its flexibility and popularity, Twitter has become the main target for spamming activities such as phishing legitimate users or spreading malicious software, which introduces new security issues and waste resources. Therefore, researchers have developed various machine-learning algorithms to reveal Twitter spam. However, as spammers have become smarter and more crafty, the characteristics of the spam tweets are varying over time making these methods inefficient to detect new spammers tricks and strategies. In addition, some of the employed methods (e.g. blacklisting) or spammer features (e.g. graph-based features) are extremely time-consuming, which hinders the ability to detect spammer activities in real-time. In this paper, we introduce a framework to deal with the volatility of the spam contents and new spamming patterns, called the spam drift. The framework combines the strength of unsupervised machine learning approach, which learns from unlabeled tweets, to retrain a real-time supervised tweet-level spam detection model in a batch mode. A set of experiments on a large-scale data set show the effectiveness of the proposed online unsupervised method in adaptively discovers and learns the patterns of new spam activities and achieve stable recall values reaching more than 95%. Although the average spam precision of our method is around 60%, the high spam recall values show the ability of our proposed method in reducing spam drift problems compared to traditional machine learning algorithms.

Cross-Modality Microblog Sentiment Prediction via Bi-Layer Multimodal Hypergraph Learning

Microblog sentiment prediction has attracted extensive research focus with wide application prospects. With the increasing proportion of multimodal tweets consisting of images, texts, and emoticons, new challenges have been raised to the existing sentiment prediction schemes. More crucially, it remains an open problem to model the dependency among multiple modalities, where one or more modalities may be missing. In this paper, we present a novel Bi-layer Multimodal Hypergraph learning (Bi-MHG) toward robust sentiment prediction of multimodal tweets to tackle the above challenges. In particular, we design a two-layer structure for the proposed Bi-MHG model: The first layer, that is, a tweet-level hypergraph, learns the tweet-feature correlation and the tweet relevance to predict the sentiments of unlabeled tweets. The second layer, that is, a feature-level hypergraph learns the relevance among different feature modalities (including the midlevel visual features in Sentibank [1] ) by leveraging prior multimodal sentiment dictionaries. These two layers are connected by sharing the relevance of multimodal features in a unified bilayer learning scheme. In such a way, Bi-MHG explicitly models the modality relevance rather than implicitly weighting multimodal features adopted in the existing Multimodal Hypergraph learning [2] . Finally, a nested alternating optimization is further proposed for Bi-MHG parameter learning. We have carried out extensive evaluations on a real-world microblog dataset crawled from Sina Weibo. For the task of multimodal sentiment prediction, superior performance is reported over several state-of-the-art and alternative approaches, which demonstrates the merits of the proposed scheme.

Statistical Features-Based Real-Time Detection of Drifted Twitter Spam

Twitter spam has become a critical problem nowadays. Recent works focus on applying machine learning techniques for Twitter spam detection, which make use of the statistical features of tweets. In our labeled tweets data set, however, we observe that the statistical properties of spam tweets vary over time, and thus, the performance of existing machine learning-based classifiers decreases. This issue is referred to as “Twitter Spam Drift”. In order to tackle this problem, we first carry out a deep analysis on the statistical features of one million spam tweets and one million non-spam tweets, and then propose a novel Lfun scheme. The proposed scheme can discover “changed” spam tweets from unlabeled tweets and incorporate them into classifier's training process. A number of experiments are performed to evaluate the proposed scheme. The results show that our proposed Lfun scheme can significantly improve the spam detection accuracy in real-world scenarios.

Enhancing Semantic Role Labeling for Tweets Using Self-Training

Semantic Role Labeling (SRL) for tweets is a meaningful task that can benefit a wide range of applications such as fine-grained information extraction and retrieval from tweets. One main challenge of the task is the lack of annotated tweets, which is required to train a statistical model. We introduce self-training to SRL, leveraging abundant unlabeled tweets to alleviate its depending on annotated tweets. A novel strategy of tweet selection is presented, ensuring the chosen tweets are both correct and informative. More specifically, the correctness is estimated according to the labeling confidences and agreement of two Conditional Random Fields based labelers, which are trained on the randomly evenly spitted labeled data; while the informativeness is in proportion to the maximum distance between the tweet and the already selected tweets. We evaluate our method on a human annotated data set and show that bootstrapping improve a baseline by 3.4% F1.