All-in-one: Multi-task Learning for Rumour Verification

Proceedings of the 27th International Conference on Computational Linguistics, pages 3402–3413 Santa Fe, New Mexico, USA, August 20-26, 2018. 3402 All-in-one: Multi-task Learning for Rumour Verification Elena Kochkina1,2, Maria Liakata1,2, Arkaitz Zubiaga1 1 University of Warwick, Coventry, United Kingdom 2 Alan Turing Institute, London, United Kingdom {E.Kochkina, M.Liakata, A.Zubiaga}@warwick.ac.uk Abstract Automatic resolution of rumours is a challenging task that can be broken down into smaller components that make up a pipeline, including rumour detection, rumour tracking and stance classification, leading to the final outcome of determining the veracity of a rumour. In previous work, these steps in the process of rumour verification have been developed as separate compo- nents where the output of one feeds into the next. We propose a multi-task learning approach that allows joint training of the main and auxiliary tasks, improving the performance of rumour verification. We examine the connection between the dataset properties and the outcomes of the multi-task learning models used. Title and Abstract in Russian Три в одном: Многозадачное Обучение для Верификации Слухов Автоматическая верификация слухов в Интернете - это сложная задача, которая мо- жет быть разбита на подзадачи, включающие в себя обнаружение слухов, отслежи- вание слухов, классификацию отношения пользователей к правдивости этих слухов и, в итоге, разрешение основной задачи - определение достоверности слуха. В преды- дущих исследованиях модели для решения каждой из подзадач разрабатывались как отдельные компоненты, где вывод модели для задачи предыдущего этапа являлся вводдом для модели задачи следующего этапа. Мы предлагаем использование много- задачного обучения для совместного обучения основной и вспомогательных задач в одной модели, что позволяет улучшить качество модели верификации слухов. Также мы рассматриваем связь между свойствами набора данных и результатами исполь- зования моделей с многозадачным обучением. 1 Introduction Social media have gained popularity as platforms that enable users to follow events and breaking news (Hu et al., 2012). However, not all information that spreads on social media during such events is accurate. The serious harm that inaccurate information can cause to society in critical situations (Lewandowsky et al., 2012) has led to an increased interest within the scientific community to develop tools to verify information from social media (Shu et al., 2017). Likewise, social media platforms them- selves, such as Facebook,1 are investing significant effort in mitigating the problems caused by misin- formation. One of the features that characterises social media is the rapid emergence and spread of new information. This leads to the circulation of rumours, claims that are unverified at the time of posting, and for which eventually evidence proving their true or false nature may come out (DiFonzo and Bordia, 2007). Therefore, a rumour resolution system needs to go through a set of steps, from detecting that a new circulating claim is a rumour, to the ultimate step of determining its veracity value. This work is licenced under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/ 1https://www.recode.net/2017/2/16/14632726/mark-zuckerberg-facebook-manifesto-fake-news-terrorism 3403 The rumour resolution process has been defined as a pipeline involving four sub-tasks (Zubiaga et al., 2018a) (see Figure 1): (1) rumour detection, determining …

Content truncated to 500 words