id stringlengths 12 14 | text stringlengths 0 19k | title stringlengths 0 183 |
|---|---|---|
1611.06322_3 | Detecting these rumours allows debunking them to prevent them from further spreading and causing harm. The further a rumour has spread, the more likely it is to be debunked by users or traditional media (Liu et. al, 2015). However, by then rumours might have already caused harm. This highlights the importance and neces... | Introduction |
1611.06322_4 | Rumour detection on social media is challenging due to the short texts, creative lexical variations and high volume of the streams. The task becomes even harder if we attempt to perform rumour detection on-the-fly, without looking into the future. We provide an effective and highly scalable approach to detect rumours i... | Introduction |
1611.06322_5 | The contributions of this paper include: | Introduction |
1611.06322_6 | Novelty based Features | Introduction |
1611.06322_7 | We introduced a new category of features for instant rumour detection that harnesses trusted resources. Unconfirmed (novel) information with respect to trusted resources is considered as an indication of rumours. | Introduction |
1611.06322_8 | Pseudo Feedback for Detection/Classification | Introduction |
1611.06322_9 | Pseudo feedback increases detection accuracy by harnessing repeated signals, without the need of retrospective operation. | Introduction |
1611.06322_10 | Before rumour detection, scientists already studied the related problem of information credibility evaluation (Castillo et. al. 2011; Richardson et. al, 2003). Recently, automated rumour detection on social media evolved into a popular research field which also relies on assessing the credibility of messages and their ... | Related Work |
1611.06322_11 | Many of these context based features originate from a study by Castillo et. al (2011), which pioneered in engineering features for credibility assessment on Twitter (Liu et. al, 2015). They observed a significant correlation between the trustworthiness of a tweet with context-based characteristics including hashtags, p... | Related Work |
1611.06322_12 | While studying the trustworthiness of tweets during crises, Mendoza et. al (2010) found that the topology of a distrustful tweet's propagation pattern differs from those of news and normal tweets. These findings along with the fact that rumours tend to more likely be questioned by responses than news paved the way for ... | Related Work |
1611.06322_13 | Recent research mainly focuses on further improving the quality of rumour detection while neglecting the increasing delay between the publication and detection of a rumour. The motivation for rumour detection lies in debunking them to prevent them from spreading and causing harm. Unfortunately, state-of-the-art systems... | Related Work |
1611.06322_14 | Therefore, researchers like Liu et. al (2015), Wu et. al (2015), Zhao et. al (2015) and Zhou et. al (2015) focus on 'early rumour-detection' while allowing a delay up to 24 hours. Their focus on latency aware rumour detection makes their approaches conceptually related to ours. Zhao et. al (1015) found clustering tweet... | Related Work |
1611.06322_15 | In contrast, we introduce new features to detect rumours as early as possible - preferably instantly, allowing them to be debunked before they spread and cause harm. | Related Work |
1611.06322_16 | Rumour detection is a challenging task, as it requires determining the truth of information (Zhao et. al, 2015). The Cambridge dictionary, defines a rumour as information of doubtful or unconfirmed truth. We rely on classification using an SVM, which is the state-of-the-art approach for novelty detection. Numerous feat... | Rumour Detection |
1611.06322_17 | We frame the Real-time Rumour Detection task as a classification problem that assesses a document's likelihood of becoming a future rumour at the time of its publication. Consequently, prediction takes place in real-time with a single pass over the data. | Problem Statement |
1611.06322_18 | More formally, we denote by $d_t$ the document that arrives from stream $S:\lbrace d_0, d_1, . . . d_n\rbrace $ at time $t$ . Upon arrival of document $d_t$ we compute its corresponding feature vector $f_{d,t}$ . Given $f_{d,t}$ and the previously obtained weigh vector $w$ we compute the rumour score $RS_{d,t} = w^T \t... | Problem Statement |
1611.06322_19 | To increase instantaneous detection performance, we compensate for the absence of future information by consulting additional data sources. In particular, we make use of news wire articles, which are considered to be of high credibility. This is reasonable as according to Petrovic et. al (2013), in the majority of case... | Novelty-based Features |
1611.06322_20 | High volume streams demand highly efficient feature computation. This applies in particular to novelty based features since they can be computationally expensive. We explore two approaches to novelty computation: one based on vector proximity, the other on kterm hashing. | Novelty Feature Construction |
1611.06322_21 | Computing novelty based on traditional vector proximity alone does not yield adequate performance due to the length discrepancy between news wire articles and social media messages. To make vector proximity applicable, we slide a term-level based window, whose length resembles the average social media message length, t... | Novelty Feature Construction |
1611.06322_22 | The second approach to compute novelty relies on kterm hashing (Wurzer et. al, 2015), a recent advance in novelty detection that improved the efficiency by an order of magnitude without sacrificing effectiveness. Kterm hashing computes novelty non-comparatively. Instead of measuring similarity between documents, a sing... | Novelty Feature Construction |
1611.06322_23 | When kterm hashing was introduced by Wurzer et. al (2015) for novelty detection on English tweets, they weighted all kterm uniformly. We found that treating all kterms as equally important, does not unlock the full potential of kterm hashing. Therefore, we additionally extract the top 10 keywords ranked by $tf.idf$ and... | Novelty Feature Construction |
1611.06322_24 | Apart from novelty based features, we also apply a range of 51 context based features. The full list of features can be found in table 6 . The focus lies on features that can be computed instantly based only on the text of a message to keep the latency of our approach to a minimum. Most of these 51 features overlap wit... | Novelty Feature Construction |
1611.06322_25 | On the arrival of a new message from a stream, all its features are computed and linearly combined using weights obtained from an SVM classifier, yielding the rumour score. We then judge rumours based on an optimal threshold strategy for the rumour score. | Novelty Feature Construction |
1611.06322_26 | In addition to novelty based features we introduce another category of features - dubbed Pseudo-Feedback (PF) feature - to boost detection performance. The feature is conceptually related to pseudo relevance feedback found in retrieval and ranking tasks in IR. The concept builds upon the idea that documents, which reve... | Pseudo Feedback |
1611.06322_27 | The PF feature describes the maximum similarity between a new document and those documents previously considered as rumour. Similarities are measured by vector proximity in term space. Conceptually, PF passes on evidence to repeated signals by increasing the rumour score of future documents if they are similar to a rec... | Pseudo Feedback |
1611.06322_28 | Training Pseudo Feedback Features | Pseudo Feedback |
1611.06322_29 | The trainings routine differs from the standard procedure, because the computation of the PF feature requires two training rounds as we require a model of all other features to identify 'pseudo' rumours. In a first training round a SVM is used to compute weights for all features in the trainings set, except the PF feat... | Pseudo Feedback |
1611.06322_30 | The previous sections introduced two new categories of features for rumour detection. Now we test their performance and impact on detection effectiveness and efficiency. In a streaming setting, documents arrive on a continual basis one at a time. We require our features to compute a rumour-score instantaneously for eac... | Experiments |
1611.06322_31 | We report accuracy to evaluate effectiveness, as is usual in the literature (Zhou et. al, 2015). Additionally we use the standard TDT evaluation procedure (Allan et. al, 2000; NIST, 2008) with the official TDT3 evaluation scripts (NIST, 2008) using standard settings. This procedure evaluates detection tasks using Detec... | Evaluation metrics |
1611.06322_32 | Rumour detection on social media is a novel research field without official data sets. Since licences agreements forbid redistribution of data, no data sets from previous publications are available. We therefore followed previous researchers like Liu et. al (2015) and Yang et. al (2012) and created our own dataset. | Data set |
1611.06322_33 | trusted resources: We randomly collected 200 news articles about broad topics commonly reported by news wires over our target time period. These range from news about celebrities and disasters to financial and political affairs as seen in table 1 . Since we operate on Chinese social media, we gathered news articles fro... | Data set |
1611.06322_34 | For our social media stream, we chose Sina Weibo, a Chinese social media service with more than 200 million active users. Micro-blogs from Sina Weibo are denoted as 'weibos'. | Data set |
1611.06322_35 | rumours: Sina Weibo offers an official rumour debunking service, operated by trained human professionals. Following Yang et. al (2012) and Zhou et. al (2015), we use this service to obtain a high quality set of 202 confirmed rumours. | Data set |
1611.06322_36 | non-rumours: We additionally gathered 202 non-rumours using the public Sina Weibo API. Three human annotators judged these weibos based on unanimous decision making to ensure that they don't contain rumours. | Data set |
1611.06322_37 | Since we operate in a streaming environment, all weibos are sorted based on their publication time-stamp. Table 3 shows a list of example for rumours found in our data set. | Data set |
1611.06322_38 | We ordered the rumours and non-rumours chronologically and divided them in half, forming a training and test set. We ensured that each of the sets consists of 50% rumours and non-rumours. This is important when effectiveness is measured by accuracy. All training and optimization use the trainings set. Performance is th... | Data set |
1611.06322_39 | To evaluate our new features for rumour detection, we compare them with two state-of-the-art early rumour detection baselines Liu et. al (2015) and Yang et. al (2012), which we re-implemented. We chose the algorithm by Yang et. al (2012), dubbed Yang, because they proposed a feature set for early detection tailored to ... | Rumour detection effectiveness |
1611.06322_40 | Table 2 compares the performance of our features with the two classifiers on the 101 rumours and 101 non-rumours of the test set, when detecting rumour instantly after their publication. The table reveals comparable accuracy for Yang and Liu at around 60%. Our observed performance of Yang matches those by Liu et. al (2... | Rumour detection effectiveness |
1611.06322_41 | When increasing the detection delay to 12 and 24 hours, all three algorithms reach comparable performance with no statistically significant difference, as seen in table 4. For our approach, none of the features are computed retrospectively, which explains why the performance does not change when increasing the detectio... | Rumour detection effectiveness |
1611.06322_42 | We group our 57 features into 7 categories shown in Table 6 and analyse their contribution using feature ablation, as seen in Table 5 . Feature ablation illustrates the importance of a feature by measuring performance, when removing it from the set of features. Novelty related features based on kterm hashing were found... | Feature analysis |
1611.06322_43 | Novelty based Features | Feature analysis |
1611.06322_44 | Novelty based features revealed the highest impact on detection performance. In particular kterms formed from the top keywords contribute the most. This is interesting, as when kterm hashing was introduced (Wurzer et. al, 2015), all kterms were considered as equally important. We found that prioritising certain kterms ... | Feature analysis |
1611.06322_45 | Interestingly, novelty based features computed by the vector similarity between weibos and news sub-documents perform slightly worse (-2% absolute). When striping all but the top tf-idf weighted terms from the news sub-documents, the hit in performance can be reduced to -1 % absolute. Kterm constructs a combined memory... | Feature analysis |
1611.06322_46 | Pseudo Feedbaack | Feature analysis |
1611.06322_47 | Features ablation revealed that pseudo feedback (PF) increased detection performance by 5.3% (relative). PF builds upon the output of the other features. High performance of the other features results in higher positive impact of PF. We want to further explore the behaviour of PF when other features perform badly in fu... | Feature analysis |
1611.06322_48 | Previous approaches to rumour detection rely on repeated signals to form propagation graphs or clustering methods. Beside causing a detection delay these methods are also blind to less popular rumours that don't go viral. In contrast, novelty based feature require only a single message enabling them to detect even the ... | Detecting unpopular rumours |
1611.06322_49 | To demonstrate the high efficiency of computing novelty and pseudo feedback features, we implement a rumour detection system and measure its throughput when applied to 100k weibos. We implement our system in C and run it using a single core on a 2.2GHz Intel Core i7-4702HQ. We measure the throughput on an idle machine ... | Efficiency and Scalability |
1611.06322_50 | We introduced two new categories of features which significantly improve instantaneous rumour detection performance. Novelty based features consider the increased presence of unconfirmed information within a message with respect to trusted sources as an indication of being a rumour. Pseudo feedback features consider me... | Conclusion |
1604.02038_0 | We propose Sentence Level Recurrent Topic Model (SLRTM), a new topic model that assumes the generation of each word within a sentence to depend on both the topic of the sentence and the whole history of its preceding words in the sentence. Different from conventional topic models that largely ignore the sequential orde... | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves |
1604.02038_1 | Statistic topic models such as Latent Dirichlet Allocation (LDA) and its variants BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 have been proven to be effective in modeling textual documents. In these models, a word token in a document is assumed to be generated by a hidden mixture model, where the hidden variables a... | Introduction |
1604.02038_2 | Most topic models take the bag-of-words assumption, in which every document is treated as an unordered set of words and the word tokens in such a document are sampled independently with each other. The bag-of-words assumption brings computational convenience, however, it sacrifices the characterization of sequential pr... | Introduction |
1604.02038_3 | There have been some works trying to solve the aforementioned problems, although still insufficiently. For example, several sentence level topic models BIBREF9 , BIBREF10 , BIBREF11 tackle the topic coherence problem by assuming all the words in a sentence to share the same topic (i.e., every sentence has only one topi... | Introduction |
1604.02038_4 | We propose Sentence Level Recurrent Topic Model (SLRTM) to tackle the limitations of the aforementioned works. In the new model, we assume the words in the same sentence to share the same topic in order to guarantee topic coherence, and we assume the generation of a word to rely on the whole history in the same sentenc... | Introduction |
1604.02038_5 | With the proposed SLRTM, we can not only model the document generation process more accurately, but also construct new natural sentences that are coherent with a given topic (we call it topic2sentence, similar to image2sentece BIBREF19 ). Topic2sentence has its huge potential for many real world tasks. For example, it ... | Introduction |
1604.02038_6 | We have conducted experiments to compare SLRTM with several strong topic model baselines on two tasks: generative model evaluation (i.e. test set perplexity) and document classification. The results on several benchmark datasets quantitatively demonstrate SLRTM's advantages in modeling documents. We further provide som... | Introduction |
1604.02038_7 | One of the most representative topic models is Latent Dirichlet Allocation BIBREF2 , in which every word in a document has its topic drawn from document level topic weights. Several variants of LDA have been developed such as hierarchical topic models BIBREF22 and supervised topic models BIBREF3 . With the recent devel... | Related Work |
1604.02038_8 | Most of the aforementioned works take the bag-of-words assumption, which might be too simple according to our discussions in the introduction. That is, it ignores both sequential dependency of words and topic coherence of words. | Related Work |
1604.02038_9 | There are some efforts trying to address the limitations of the bag-of-words assumption. For example, in BIBREF27 , both semantic (i.e., related with topics) and syntactic properties of words were modeled. After that, a hidden Markov transition model for topics was proposed BIBREF9 , in which all the words in a sentenc... | Related Work |
1604.02038_10 | Another line of research related to our model is Recurrent Neural Network (RNN), especially some recently developed effective RNN models such as Long Short Term Memory BIBREF17 and Gated Recurrent Unit BIBREF18 . These new RNN models characterize long range dependencies for a sequence, and has been widely adopted in se... | Related Work |
1604.02038_11 | In this section, we describe the proposed Sentence Level Recurrent Topic Model (SLRTM). First of all, we list three important design factors in SLRTM as below. | Sentence Level Recurrent Topic Model |
1604.02038_12 | With the three points in mind, let us introduce the detailed generative process of SLRTM, as well as the stochastic variational inference and learning algorithm for SLRTM in the following subsections. | Sentence Level Recurrent Topic Model |
1604.02038_13 | Suppose we have INLINEFORM0 topics, INLINEFORM1 words contained in dictionary INLINEFORM2 , and INLINEFORM3 documents INLINEFORM4 . For any document INLINEFORM5 , it is composed of INLINEFORM6 sentences and its INLINEFORM7 th sentence INLINEFORM8 consists of INLINEFORM9 words. Similar to LDA, we assume there is a INLIN... | The generative process |
1604.02038_14 | Sample the multinomial parameter INLINEFORM0 from INLINEFORM1 ; | The generative process |
1604.02038_15 | For the INLINEFORM0 th sentence of document INLINEFORM1 INLINEFORM2 , INLINEFORM3 , where INLINEFORM4 is the INLINEFORM5 th word for INLINEFORM6 : | The generative process |
1604.02038_16 | Draw the topic index INLINEFORM0 of this sentence from INLINEFORM1 ; | The generative process |
1604.02038_17 | For INLINEFORM0 : | The generative process |
1604.02038_18 | Compute LSTM hidden state INLINEFORM0 ; | The generative process |
1604.02038_19 | INLINEFORM0 , draw INLINEFORM1 from DISPLAYFORM0 | The generative process |
1604.02038_20 | The generative process | |
1604.02038_21 | Here we use bold characters to denote the distributed representations for the corresponding items. For example, INLINEFORM0 and INLINEFORM1 denote the embeddings for word INLINEFORM2 and topic INLINEFORM3 , respectively. INLINEFORM4 is a zero vector and INLINEFORM5 is a fake starting word. Function INLINEFORM6 is the L... | The generative process |
1604.02038_22 | where INLINEFORM0 , INLINEFORM1 denotes the output embedding for word INLINEFORM2 . INLINEFORM3 are feedforward weight matrices and INLINEFORM4 is the bias vector. | The generative process |
1604.02038_23 | Then the probability of observing document INLINEFORM0 can be written as: DISPLAYFORM0 | The generative process |
1604.02038_24 | where INLINEFORM0 is the probability of generating sentence INLINEFORM1 under topic INLINEFORM2 , and it is decomposed through the probability chain rule; INLINEFORM3 is specified in equation ( EQREF11 ) and ( EQREF12 ); INLINEFORM4 represents all the model parameters, including the distributed representations for all ... | The generative process |
1604.02038_25 | To sum up, we use Figure FIGREF14 to illustrate the generative process of SLRTM, from which we can see that in SLRTM, the historical words and topic of the sentence jointly affect the LSTM hidden state and the next word. | The generative process |
1604.02038_26 | As the computation of the true posterior of hidden variables in equation ( EQREF13 ) is untractable, we adopt mean field variational inference to approximate it. Particularly, we use multinomial distribution INLINEFORM0 and Dirichlet distribution INLINEFORM1 as the variational distribution for the hidden variables INLI... | Stochastic Variational Inference and Learning |
1604.02038_27 | where INLINEFORM0 is the true distribution for corresponding variables. | Stochastic Variational Inference and Learning |
1604.02038_28 | The introduction of LSTM-RNN makes the optimization of ( EQREF16 ) computationally expensive, since we need to update both the model parameters INLINEFORM0 and variational parameters INLINEFORM1 after scanning the whole corpus. Considering that mini-batch (containing several sentences) inference and training are necess... | Stochastic Variational Inference and Learning |
1604.02038_29 | [ht] Stochastic Variational EM for SLRTM Input: document INLINEFORM0 , variation parameters INLINEFORM1 , and model weights INLINEFORM2 . every sentence minibatch INLINEFORM3 in INLINEFORM4 INLINEFORM5 E-Step: INLINEFORM6 INLINEFORM7 , i.e., every topic index: Obtain INLINEFORM8 by LSTM forward pass. INLINEFORM9 DISPLA... | Stochastic Variational Inference and Learning |
1604.02038_30 | convergence Collect variational parameters INLINEFORM0 . M-Step: Compute the gradient INLINEFORM1 by LSTM backward pass. Use INLINEFORM2 to obtain INLINEFORM3 by stochastic gradient descent methods such as Adagrad BIBREF30 . In Algorithm SECREF15 , INLINEFORM4 is the digamma function. Equation ( EQREF18 ) guarantees t... | Stochastic Variational Inference and Learning |
1604.02038_31 | We report our experimental results in this section. Our experiments include two parts: (1) quantitative experiments, including a generative document evaluation task and a document classification task, on two datasets; (2) qualitative inspection, including the examination of the sentences generated under each topic, in ... | Experiments |
1604.02038_32 | We compare SLRTM with several state-of-the-art topic models on two tasks: generative document evaluation and document classification. The former task is to investigate the generation capability of the models, while the latter is to show the representation ability of the models. | Quantitative Results |
1604.02038_33 | We base our experiments on two benchmark datasets: | Quantitative Results |
1604.02038_34 | 20Newsgroup, which contains 18,845 emails categorized into 20 different topical groups such as religion, politics, and sports. The dataset is originally partitioned into 11,314 training documents and 7,531 test documents. | Quantitative Results |
1604.02038_35 | Wiki10+ BIBREF31 , which contains Web documents from Wikipedia, each of which is associated with several tags such as philosophy, software, and music. Following BIBREF25 , we kept the most frequent 25 tags and removed those documents without any of these tags, forming a training set and a test set with 11,164 and 6,161... | Quantitative Results |
1604.02038_36 | We followed the practice in many previous works and removed infrequent words. After that, the dictionary contains about INLINEFORM0 unique words for 20Newsgroup and INLINEFORM1 for Wiki10+. We adopted the NLTK sentence tokenizer to split the datasets into sentences if sentence boundaries are needed. | Quantitative Results |
1604.02038_37 | The following baselines were used in our experiments: | Quantitative Results |
1604.02038_38 | LDA BIBREF2 . LDA is the classic topic model, and we used GibbsLDA++ for its implementation. | Quantitative Results |
1604.02038_39 | Doc-NADE BIBREF24 . Doc-NADE is a representative neural network based topic model. We used the open-source code provided by the authors. | Quantitative Results |
1604.02038_40 | HTMM BIBREF9 . HTMM models consider the sentence level Markov transitions. Similar to Doc-NADE, the implementation was provided by the authors. | Quantitative Results |
1604.02038_41 | GMNTM BIBREF12 . GMNTM considers models the order of words within a sentence by a feedforward neural network. We implemented GMNTM according the descriptions in their papers by our own. | Quantitative Results |
1604.02038_42 | For SLRTM, we implemented it in C++ using Eigen and Intel MKL. For the sake of fairness, similar to BIBREF12 , we set the word embedding size, topic embedding size, and LSTM hidden layer size to be 128, 128, and 600 respectively. In the experiment, we tested the performances of SLRTM and the baselines with respect to d... | Quantitative Results |
1604.02038_43 | We measure the performances of different topic models according to the perplexity per word on the test set, defined as INLINEFORM0 , where INLINEFORM1 is the number of words in document INLINEFORM2 . The experimental results are summarized in Table TABREF33 . Based on the table, we have the following discussions: | Quantitative Results |
1604.02038_44 | Our proposed SLRTM consistently outperforms the baseline models by significant margins, showing its outstanding ability in modelling the generative process of documents. In fact, as tested in our further verifications, the perplexity of SLRTM is close to that of standard LSTM language model, with a small gap of about 1... | Quantitative Results |
1604.02038_45 | Models that consider sequential property within sentences (i.e., GMNTM and SLRTM) are generally better than other models, which verifies the importance of words' sequential information. Furthermore, LSTM-RNN is much better in modelling such a sequential dependency than standard feed-forward networks with fixed words wi... | Quantitative Results |
1604.02038_46 | In this experiment, we fed the document vectors (e.g., the INLINEFORM0 values in SLRTM) learnt by different topic models to supervised classifiers, to compare their representation power. For 20Newsgroup, we used the multi-class logistic regression classifier and used accuracy as the evaluation criterion. For Wiki10+, s... | Quantitative Results |
1604.02038_47 | All the classification results are shown in Table TABREF37 . From the table, we can see that SLRTM is the best model under each setting on both datasets. We can further find that the embedding based methods (Doc-NADE, GMNTM and SLRTM) generate better document representations than other models, demonstrating the represe... | Quantitative Results |
1604.02038_48 | In this subsection, we demonstrate the capability of SLRTM in generating reasonable and understandable sentences given particular topics. In the experiment, we trained a larger SLRTM with 128 topics on a randomly sampled INLINEFORM0 Wikipedia documents in the year of 2010 with average 275 words per document. The dictio... | Qualitative Results |
1604.02038_49 | We used two different mechanisms in sentence generating. The first mechanism is random sampling new word INLINEFORM0 at every time step INLINEFORM1 from the probability distribution defined in equation ( EQREF13 ). The second is dynamic programming based beam search BIBREF19 , which seeks to generate sentences by globa... | Qualitative Results |
1604.02038_50 | The generating results are shown in Table TABREF40 . In the table, the sentences generated by random sampling and beam search are shown in the second and the third columns respectively. In the fourth column, we show the most representative words for each topics generated by SLRTM. For this purpose, we constrained the m... | Qualitative Results |
1604.02038_51 | From the table we have the following observations: | Qualitative Results |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.