Presentation is loading. Please wait.

Presentation is loading. Please wait.

Stance Classification for Fact-Checking

Similar presentations


Presentation on theme: "Stance Classification for Fact-Checking"— Presentation transcript:

1 Stance Classification for Fact-Checking
Lecture: Web Science Luca Brandt

2 Table of Content Introduction Motivation Fact-Checking Process
Paper 1 – Fake News Challenge Paper 2 – Relevant Document Discovery for Fact-Checking Articles Future Work Luca Brandt Stance Classification Web Science 2019

3 Introduction What are Fake News ? Problem with Fake News
Propaganda (not objective information) Deliberate disinformation/hoaxes Reporters paying sources for stories Made up stories Problem with Fake News Spreading misinformation Reduces trust in news media Manipulating society Luca Brandt Stance Classification Web Science 2019

4 Introduction What is Fact-Checking? Claim
Process to check the veracity and correctness of a claim/statement In non-fictional texts A classic journalism task Claim Statement made by a politician Story published by another journalist Rumor on Social Media Etc. Luca Brandt Stance Classification Web Science 2019

5 Example Claim Luca Brandt Stance Classification Web Science 2019

6 Motivation Why Fact-Checking? Fact-Checking and AI
Goal is to provide a verdict if a claim is true, false or mixed Fight misinformation Identify Fake News Providing context to users to understand information better Fact-Checking and AI Automatically detect Fake News Automatically gather documents relevant to a claim We want to automate the detection of fake news because it is Time consuming For 1 article to be fact checked 1 person needs approx 1 day Luca Brandt Stance Classification Web Science 2019

7 Fact-Checking Process
Given a claim/statement Find documents relevant to claim Classic Information Retrieval task Understand the stance of relevant documents Stance Classification, Classification Problem Give verdict if claim is true or false Classification Problem Understand stance: this is Stance Detection -> given a document and a claim -> what stance has the document regarding the claim is it true or false or mixed Luca Brandt Stance Classification Web Science 2019

8 Fake News Challenge (FNC-1)
Foster development of AI tech to detect Fake News 50 teams participated from industry and academia Task: Stance Detection of an entire document Learn classifier f: (document, headline) -> stance {AGR,DSG,DSC,UNR} Dataset: 300 Topics – claims with 5-20 documents each Every document summarized to a headline Each document matched with every headline to generate dataset 2017 Stance of entire document -> earlier stance classification of a single sentence Luca Brandt Stance Classification Web Science 2019

9 A Retrospective Analysis of the Fake News Challenge Stance Detection Task
By Hanselowski et al. 13. June 2018 Main Contributions: First summarizing and analyzing paper for the FNC-1 Reproduction of results of top-3 performers Proposed a new evaluation metric Proposed a new model Year of publication Luca Brandt Stance Classification Web Science 2019

10 Top-3 Performers 1. TalosComb 2. Athene 3. UCL Machine Reading (UCLMR)
Weighted average model of deep conv net and gradient-boosted decision tree TalosCNN uses pre-trained word2vec embeddings TalosTree based on word count, TF-IDF, sentiment, and word2vec embeddings 2. Athene Multilayer perceptron (MLP) with 6 hidden layers, with handcrafted features Unigrams, cosine similarity, topic models 3. UCL Machine Reading (UCLMR) MLP as well but only 1 layer Term frequency vectors of 5000 most frequent unigrams Cosine Similarity between TF-IDF vectors of headline and doc TalosCNN contains several Conv layers and three fully contected layers + softmax to obtain probabilities UCL Cosine similarity of word embeddings of nouns and verbs in headline and document Luca Brandt Stance Classification Web Science 2019

11 Problem with the metric and dataset
Hierarchical metric .25 points if classified correctly as related {AGR, DSG, DSC} or unrelated {UNR} .75 points if classified correctly as AGR, DSG, or DSC But related class is imbalanced Not difficult to predict related or unrelated (best systems reach 99% UNR) Correctly predicting related vs unrelated and always picking DSC would achieve a FNC-1 score of > better than the winner .25 points if document classified correctly given a headline Dataset So it favors models which correctly predict majority classes, but it is more interesting to correctly predict the minority classes AGR or DSG in the case of factchecking! Luca Brandt Stance Classification Web Science 2019

12 Their model and metric F1m metric StackLSTM
Class-wise F1 scores and macro average to F1m score F1 = Not affected by large size of majority class Naive approach predicting UNR and always DSC -> F1m = .444 StackLSTM Combines best features of their feature test Concatenate GloVe Wordembeddings fed through 2 stacked LSTMs Understand the meaning of whole sentence Hidden state of LSTMs fed through 3 layers NN Softmax to obtain probabilities Combines best features of their test: they tested the features of the top 3 models, how much they contribute and what are the most promising features GloVe Embeddings highlight Concatenate up tp 100 tokens of the 50-dim glove embedd. then fed through the LSTMs With this technique they produce sequences of words which hold more semantic meaning than uni and bi grams and can capture more complex patterns Luca Brandt Stance Classification Web Science 2019

13 Reproduction of Results
Reproduction of all results successfull without significant deviation TalosComb is combination of CNN and Tree so they included all three FNC is originall metric F1m is their proposed metric Class wise F1 scores Regarding FNC score Athene won But Talos Tree actually performs better than TalosComb stackLSTM is best but not significant But is significantly better than every other model in terms of DSG class the minority class -> which is important for fact checking Luca Brandt Stance Classification Web Science 2019

14 Pros & Cons Pros: Cons: First paper to summarize and analyze the FNC-1
First paper to reproduce the results Proposed a better metric Proposed a new model better than the state of the art Cons: Proposed new model still has low accuracy of Disagreeing class Luca Brandt Stance Classification Web Science 2019

15 Relevant Document Discovery for Fact-Checking Articles
Paper by Wang et al. 23. April 2018 Main Contributions: End-to-end system for relevant document discovery for fact-cheking articles Better than state of the art stance classification Better than state of the art relevance classification Better state of the art Luca Brandt Stance Classification Web Science 2019

16 Fact-Checking Articles
Adopted Schema.org ClaimReview Markup Provides structure of an article Key Fields on top of content: Claim Claimant Verdict Structured fields cannot provide documents relevant to claim Identifying claim relevant documents extremely usefull Luca Brandt Stance Classification Web Science 2019

17 Fact Checking Article & Claim Relevant Doc
Luca Brandt Stance Classification Web Science 2019

18 Overview of their system
Start with a fact checking article Based on the factchecking article they try to find a set of related articles Binary classifier: check whether a related document is relevant or not Use another classifier to predict the stance of relevant documents wrt the fact checking article and its claim Difference to the systems of other paper Here they start of a fact-checking article In the other paper they started with a document headline pair. In this paper they do a pre classification to eliminate the unrelevant documents In the other paper they had a lot of unrelated instances. Luca Brandt Stance Classification Web Science 2019

19 Candidate Generation Via Navigation Via Search with Google
Outgoing links from fact-checking article But most of them not relevant Via Search with Google Key challenge: generate the right set of queries Texts from title and claim Title and claim text transformed with entity annotations Click graph queries Combining both generating about 2400 related documents Finding the related documents Because most of navigational sources not relevant -> generate candidates through search Entity annotations: extract important information from title and claim -> only entities Click graph queries: queries which led a user to click on the fact checking article they collect up to 50 queries Luca Brandt Stance Classification Web Science 2019

20 Relevance Classification
Classifier M: (f,d) -> {relevant, irrelevant} Features: Building confidence vectors of entities Cosine similarity between confidence vectors of Claim and text/sentence/paragraph of related doc Sentence of fact-checking article and sentence of related doc And whole documents Publication date Gradient Boosted Decision Tree Combines all features – predicts relevant or irrelevant Confidence vectors: how often a word is found in fact-checking article and the entire related doc Luca Brandt Stance Classification Web Science 2019

21 Stance Classification
Build model M: (f, d) -> {contradict, support} Similarity not good for stance classification Find key contradicting patterns in context similar to the claim Collected 3.1k (claim, contradicting statement) pairs Built 900-dim lexicon from uni- and bi-grams with grams of highest probability Uni-grams: hoax, fake, purportedly, rumor Bi-grams: made-up, fact check, not true, no evidence Classifier input: fact-checking article and relevant doc Crowd sourcing labeling: found out 80% of relevant docs are contradict or support They say number of neither/discuss articles are small in the fact checking domain so they dont lay focus on this. Difference to the systems of the other paper There was the class Discussing which is left out here Two texts can have much lexical overlap but a single not in the sentence can change the whole meaning Where the contradicting statement was mostly the headline because usually the headline is a contradicting statement Luca Brandt Stance Classification Web Science 2019

22 Stance Classification
from relevant doc use title, headline, text and prune away text whose similarity is smaller than a threshold Concatenate text with one sentence before and after the text-> key components Extract uni- and bi-grams of key components -> final feature vector Using a Gradient boosted decision tree for prediction Luca Brandt Stance Classification Web Science 2019

23 Results of their model Interesting their model vs the FNC Winner.
Here original FNC TalosComb was evaluated on their dataset -> performed poorly And then FNC TalosComb was trained and evaluated on their dataset, but only the gradient boosted decision tree, because TalosCNN model was not trainable -> performed better but still their model was better. But you have to pay attention to the results: 91% accuracy in stance classification seems to good to be true, especially when you take the 81% accuracy of the relevance classification into account which is the pre processing for the stance classification They evaluated each part of their system in isolation. So the approx 20% error of the relevance classification is not playing a role in the stance classification If they would have evaluated the whole system the score for the stance classification would have been lower.

24 Pros & Cons Pros: New state of the art stance classification algorithm Proposed whole end-to-end system for relevant document discovery Cons: Not providing their dataset Not providing the distribution of the dataset Not providing the per class scores Evaluation in isolation Ignored the Discussing class Distribution not provided Per class scores Isolation evaluation Discussing class still exists Luca Brandt Stance Classification Web Science 2019

25 Conclusion What are Fake News What is Fact-Checking
Fake News Challenge Top-3 Performers Stack LSTM End-to-End System for Relevant Document Discovery The Disagreeing Class is not really well predicted What we showed Luca Brandt Stance Classification Web Science 2019

26 Future Work All features text based
Including non-textual data as features Videos -> Image & Speech Recognition Social Media Pages -> embedded image/graphic information Develop ML techniques with deeper semantic understanding Not relying on lexical features Disagreeing class has to be predicted better Need to higher accuracy on disagree class Luca Brandt Stance Classification Web Science 2019

27 Thank you for your attention
Any Questions?

28 Dataset of Paper 2 Unlabled Corpus: fact-checking articles by highly reputable fact- checkers Relevance-labeled corpus: using their candidate generation algorithm related documents per fact-checking article 33.5M With crowd-workers and balancing for positive and negative examples a total of (claim, doc) pairs Crowd-working question: does this doc address the claim? Stance-labeled corpus: randomly sampled 1200 of the positive instances of relevance-labled corpus and crowdsourced -> support, contradict, neither, can‘t tell For 12% workers couldn‘t agree -> removed Manual Corpus: to measure candidate generation, randomly sampled 450 fact-checking articles and let crowdworkers search for candidates Luca Brandt Stance Classification Web Science 2019

29 Features Description Bag of Words : Topic Models:
1- and 2-Grams with 5000 token vocabulary for headline and doc Added a negation flag „_NEG“ as prefix to every word between special negation words like „not“, „never“, „no´“ until next punctuation mark Topic Models: Non-negative matrix factorization Latent semantic indexing Latent dirichlet allocation Similarity between topic models and headlines and bodies Luca Brandt Stance Classification Web Science 2019

30 Feature Description II
Lexicon-based features Based on lexicons which hold the sentiment/polarity for each word Computed seperately for headline and body Count positive and negative polarized words -> features Find maximum positive and negative polarity -> features Last word with negative or positive polarity -> feature Refuting words list („fake“, „hoax“) -> features Concatenating all features from above Readability features Measured with different metrics (e.g. SMOG, Flesch-Kincaid, Gunning fog) Luca Brandt Stance Classification Web Science 2019


Download ppt "Stance Classification for Fact-Checking"

Similar presentations


Ads by Google