Stance classification and Diffusion Modelling

Transcription

1 Dr. Srijith P. K. CSE, IIT Hyderabad

2 Outline Stance Classification 1 Stance Classification 2

3 Stance Classification in Twitter Rumour Stance Classification Classify tweets as supporting, denying, questioning, or commenting an event. Useful for rumour truthfullness classification.

4 Stance Classification in Twitter Time Sensitive Sequence Classification Tweets form a conversation structure with reply tweets. Tweet classification as sequence labelling problem. Each tweet is associated with the time of occurrence.

5 Point Processes in Twitter Stance Classification Twitter data containing information tweet time, text, meme category and user {d n = (t n, W n, m n, i n)} N n=1. Models the spiky behaviour typically observed in social networks Self exciting point process with intensity for user λ i (t) = µ i + t l <t I(i l == i)α i κ(t t l ) Base intensity influence of past tweets Past tweets influence future tweets but decays exponentially over time κ(t t l ) = ω exp( ω(t t l ))

6 Point Processes in Twitter Stance Classification Twitter data containing information tweet time, text, meme category and user {d n = (t n, W n, m n, i n)} N n=1. Multivariate Hawkes models the influence of other users. λ i (t) = µ i + t l <t α i l,i κ(t t l ) Joint modelling of users and memes with multivariate Hawkes. λ i,m (t) = µ i γ m + t l <t I(m l == m)α il,i κ(t t l )

7 Time Sensitive Sequence Classification [Lukasik et al., 2016] Twitter data containing tweet time, text and label {d n = (t n, W n, m n, y n)} N n=1. Multivariate for stance classification Intensity modelled over labels λ y,m(t) = µ y + t l <t I(m l == m)α yl,y κ(t t l ) Classification of tweet depends on the textual Content p(w n y n) = V v=1 βwnv y nv. Likelihood of a tweet belonging to a class is proportional to p(w n y n) λ yn,m n (t n) emission likelihood transition likelihood Generative non-markovian model : Past tweet labels influence future tweet labels Generalizes Multinomial, Naive Bayes, hidden markov models.

8 Time Sensitive Sequence Classification Learning and Prediction Likelihood [ N N ] n=1 p(wn yn) n=1 λyn,mn (t n) p(e T ) Learning by maximum likelihood approach. Y M T y=1 m=1 0 λy,m(s)ds + N n=1 log λyn,mn (tn) + N V n=1 v=1 Wnv log βynv. Prediction of labels uses a greedy approach Datasets Dataset Tweets Supporting Denying Questioning Commenting Ottawa shooting Ferguson riots Charlie Hebdo Sydney siege

9 Time Sensitive Sequence Classification (experimental results)

10 Time Sensitive Sequence Classification (experimental results)

11 Deep learning for stance classification Convolutional Neural Network for Text Documents/sentence represented as a matrix. Row corresponds to vector representation of a word (word embeddings (word2vec/glove)) Filters slide over full rows of the matrix, width of filter same as embedding dimension.

12 Deep learning for stance classification Convolutional Neural Network for stance classification [chen et al., 2017] Learn word embedding as well as Glove representation. Uses different sized filters with number of filters of same size being 128.

13 Deep learning for stance classification Recurrent Neural Network for stance classification Use word embedding from Glove representation. Unroll over words in a tweet and predict the label.

14 Deep learning for stance classification Neural Network for stance classification deep learning model is the best method in terms of F1 score.

15 Deep learning for stance classification CNN and LSTM combination [amir et al., 2017] deep learning model is the best method in terms of F1 score.

16 Stance Classification Model evolution of memes over time. Model the behaviours of users, how they influence each other. Predict their popularity of memes

17 Predicting Rumour Popularity Stance Classification Predict rumour popularity measured as number of tweets in future time intervals Motivation: Assist officials and journalists with debunking rumours.

18 for Twitter [Srijith et al., 2017] Twitter contain information on conversation structure, users, and network. Model Hawkes process to consider Twitter Information.

19 for Twitter Stance Classification HPconversation model considering conversational structure HPconversation : Models conversational structure (spontaneous/replyto tweet) λ in,m n (t) = µ in γ mn Z nn + n l=1 I(m l == m n)z ln α il,i n κ(t t l ) Avoids performing summation over all previous terms

20 for Twitter Stance Classification HPdecomposition HPdecomposition : Decompose matrix to lower rank non-negative components. λ in,m n (t) = µ in γ mn Z nn + n l=1 I(m l == m n)z ln [I S ] il,i n κ(t t l ) Prevents overfitting by learning a reduced number of parameters.

21 for Twitter Stance Classification HPuserfeatures model considering user features HPuserfeatures : Consider user features by parameterizing influence and susceptibility matrices λ in,mn (t) = µ in γ mn Z nn + n l=1 I(m l == m n)z ln σ([x il I ] [x in S] )κ(t t l ). Learns the features determining influence and susceptibility ( (φlis +1)(φ Links ratio depends on number of followers and followees. log in +1) 2 φ out +1 ).

22 for Twitter Stance Classification HPconnection model considering connection information HPconnection : Consider network information by selectively regularizing matrix. Users which are not connected are less influential. HPregularization : Regularize matrix entries using l 2 norm.

23 for Twitter Stance Classification Learning Complete Likelihood is given by N n=1 λ i n,m n (t n) exp( R M T i=1 m=1 0 λ i,m(s)ds) inst. probabilities over all tweets survival probability Parameters are learnt by maximizing regularized log-likelihood. l(µ, γ, α, ω) = N n=1 log λ i n,m n (t n) R M T i=1 m=1 0 λ i,m(s)ds. l(µ, γ, α, ω) = N N n 1 Z n,n log(µ in γ mn ) + I(m l = m n)z l,n log α il,i n κ(t n t l ) n=1 R M T µ i γ m i=1 m=1 i=1 l=1 n=1 l=1 R N α il,i K(T t l )

24 for Twitter Stance Classification Prediction and Evaluation Prediction by modified thinning algorithm. Similar to rejection sampling Consider an upper bound λ λ(t) in [s, u] Generate a sample from homogenous Poisson process with rate λ Accept based on the ratio λ(t) λ Upper bound λ easy to obtain for! Evaluation using aligned mean squared error and number of predictions

25 for Twitter : Experimental Results (Synthetic)

26 for Twitter : Experimental Results (Synthetic)

27 for Twitter : Experimental Results (Ferguson)

28 Recurrent point process [du et al., 2017] Recurrent point process Hawkes process assumes that the influences from past events are linearly additive True relationship is not known Recurrent neural networks helps to learn the non-linear relations ships

29 Recurrent point process [du et al., 2017]

30 Recurrent point process [du et al., 2017]

31 References Srijith, P., Lukasik, M., Cohn, T., and Bontcheva, K. (2017). Longitudinal modeling of social media with hawkes process based on users and networks. In International Conference on Advances in Social Networks Analysis and Mining. Lukasik, M., Srijith, P., Cohn, T., and Bontcheva, K. (2016). Hawkes processes for continuous time sequence classification. In Association of Computational Linguistics. Yi-Chin Chen, Zhao-Yand Liu, and Hung-Yu Kao. (2017). Convolutional Neural Networks for Stance Detection and Rumor Verification In Proceedings of SemEval. ACL.. N. Du, H. Dai, R. Trivedi, U. Upadhyay, M. Gomez-Rodriguez, and L. Song. ( 2016). Recurrent Marked Temporal Point Process: Embedding Event History to Vector. In KDD.. Amir Pouran Ben Veyseh, Javid Ebrahimi, Dejing Dou, and Daniel Lowd. ( 2017). A Temporal Attentional Model for Rumor Stance Classification In CIKM.

32 Thank you Dr. P.K. Srijith Assistant Professor Computer Science and Engineering Indian Institute of Technology Hyderabad