Centre for Cyberhate Research & Policy: Real-Time Scalable Methods & Infrastructure for Modelling the Spread of Cyberhate on Social Media
Lead Research Organisation:
Cardiff University
Department Name: Sch of Social Sciences
Abstract
The UK Government's Hate Crime Action Plan (Home Office 2016) stresses the need to tackle hate speech on social media by bringing together policymakers with academics to improve the analysis and understanding of the patterns and drivers of cyberhate and how these can be addressed. Furthermore, the recent Home Affairs Select Committee Inquiry (2016) 'Hate Crime and its Violent Consequences' highlighted the role of social media in the propagation of hate speech (on which the proposers were invited to provide evidence). This proposal acknowledges the migration of hate to social media is non-trivial, and that empirically we know very little about the utility of Web based forms data for measuring online hate speech and counter hate speech at scale and in real-time. This became particularly apparent following the referendum on the UK's future in the European Union, where an inability to classify and monitor hate speech and counter speech on social media in near-real-time and at scale hindered the use of these new forms of data in policy decision making in the area of hate crime. It was months later that small-scale grey literature emerged providing a 'snap-shot' of the problem (Awan & Zempi 2016, Miller et al. 2016). In partnership with the UK Head of the Cross-Government Hate Crime Programme at the Department for Communities and Local Government (DCLG), and the London Mayor's Office for Policing and Crime's (MOPAC) new Online Hate Crime Hub, the proposed project will co-produce evidence on how social media data, harnessed by new Social Data Science methods and scalable infrastructure, can inform policy decision making. We will achieve this by taking the social media reaction to the referendum on the UK's future in the European Union as a demonstration study, and will co-develop with the Policy CI transformational New Forms of Data Capability contributions including: (i) semi-automated methods that monitor the production and spread of cyberhate around the case study and beyond; (ii) complementary methods to study and test the effectiveness of counter speech in reducing the propagation of cyberhate, and (iii) a technical system that can support real time analysis of hate and counter speech on social media at scale following 'trigger events', integrated into existing policy evidence-based decision-making processes. The system, by estimating the propagation of cyberhate interactions within social media using machine learning techniques and statistical models, will assist policymakers in identifying areas that require policy attention and better targeted interventions in the field of online hate and antagonistic content.
Planned Impact
In line with the drive behind the call, this project will co-produce a strong evidence base on the utility of social media data to inform policy development, intervention and decision making. The project will provide a case study that will demonstrate how these data, when effectively and efficiently collected, transformed and repurposed using Social Data Science tools and methods, can have a transformative impact on how governments work to address contemporary pressing social problems. We have selected cyberhate in the aftermath of the referendum on the UK's future in the EU as a case study for understanding the relationship between social media data and policymaking.
We will work closely with the Policy CI, the UK Head of the Cross-Government Hate Crime Programme at the Department for Communities and Local Government, and the London Mayor's Office for Policing and Crime Online Hate Crime Hub, to co-produce an evidence base on the utility of social media data for policy and decision making. We will achieve this by:
--Involving the UK Head of the Cross-Government Hate Crime Programme and the MOPAC Online Hate Crime Hub in the design, testing, analysis and implementation phases of the project, to ensure maximum buy-in at a policy level
--Running requirements gathering workshops with policymakers for tool and system development
--Testing the system developed in WP6 in a policy environment and writing lessons-learned report
--Conducting post-hoc interviews with policymakers to inform an ESRC Policy Evidence Briefing and an Ethics Guide for Policymakers
--Providing free access to new Lab social media hate and counter speech classification tools for not-for-profit use
We will work closely with the Policy CI, the UK Head of the Cross-Government Hate Crime Programme at the Department for Communities and Local Government, and the London Mayor's Office for Policing and Crime Online Hate Crime Hub, to co-produce an evidence base on the utility of social media data for policy and decision making. We will achieve this by:
--Involving the UK Head of the Cross-Government Hate Crime Programme and the MOPAC Online Hate Crime Hub in the design, testing, analysis and implementation phases of the project, to ensure maximum buy-in at a policy level
--Running requirements gathering workshops with policymakers for tool and system development
--Testing the system developed in WP6 in a policy environment and writing lessons-learned report
--Conducting post-hoc interviews with policymakers to inform an ESRC Policy Evidence Briefing and an Ethics Guide for Policymakers
--Providing free access to new Lab social media hate and counter speech classification tools for not-for-profit use
Publications
Alorainy W
(2019)
"The Enemy Among Us" Detecting Cyber Hate Speech with Threats-based Othering Language Embeddings
in ACM Transactions on the Web
Alorainy W
(2022)
Disrupting networks of hate: characterising hateful networks and removing critical nodes
in Social Network Analysis and Mining
Bérubé M
(2020)
Social media forensics applied to assessment of post-critical incident social reaction: The case of the 2017 Manchester Arena terrorist attack.
in Forensic science international
Javed A
(2020)
Emotions Behind Drive-by Download Propagation on Twitter
in ACM Transactions on the Web
Liu H
(2019)
A Fuzzy Approach to Text Classification With Two-Stage Training for Ambiguous Instances
in IEEE Transactions on Computational Social Systems
Ozalp, S.
(2020)
Antisemitism on Twitter: Collective efficacy and the role of community organisations in challenging online hate speech
in Social Media + Society
Williams M
(2017)
The Ethics of Online Research
Williams M
(2019)
Hate in the Machine: Anti-Black and Anti-Muslim Social Media Posts as Predictors of Offline Racially and Religiously Aggravated Crime
in The British Journal of Criminology
| Description | HateLab was originally funded under the ESRC Civil Society and Social Media call in 2017. The grant enabled the research team, which included a non-academic policy CI, to demonstrate how Digital Footprint Data, in the form of social media, could provide insights into the nature and dynamics of online hate speech and it how might be effectively marshalled to benefit policy and operational decision making. A key output is the HateLab Platform, a first of a kind Digital Footprint Data (DFD) aggregation cloud-based resource that visualises the classifications of social media hate speech based on state-of-the art artificial intelligence developed in HateLab. The Platform was implemented within HMG's National Cyber Hate Crime Hub in Q2 2019. The Hub was established by the Home Secretary in 2017 to tackle online forms of hate crime, that increased dramatically in the aftermath of the 2016 referendum vote on the future of the UK in the EU. It acts as the point of contact for all victims of online hate crime, and produces intelligence reports for police, senior civil servants and MPs. It was agreed with the Policy CI, Paul Giannasi, National Police Lead for Hate Crime, that the Platform be evaluated in two live police operations in the Hub: i) monitoring hate speech around the exit of the UK from the EU; and ii) monitoring hate speech targeting MPs. The Platform and HateLab insights allowed the Hub to better understand the dynamics of hate speech propagation, leading to improved response times, better support for victims and more effective allocation of resources. Based on this early-stage impact, the HateLab project was submitted as a Cardiff University REF2021 case study (UoA 21). In 2021, the Director, Professor Williams, published a trade non-fiction book, The Science of Hate with Faber & Faber, that draws heavily on HateLab research. More recently, VLOPs (TikTok, Twitter, Meta) have engaged with HateLab to inform the development of their community guidelines in relation to protected groups and hate speech counter-measures. Williams has recently taken up a formal advisory role at TikTok. With a within-award grant, the Platform and associated data services were rolled out to additional users in a broader pilot: Welsh Government, Galop (LGBTQ+ anti-violence charity), and South Wales Police, and later to the Home Office, The Football Association of Wales, Kick it Out, EE, BT, Saatchi & Saatchi and Deutsche Telekom. Impacts include monitoring and countering online hate speech related to: • The Brexit process • The abuse of Members of Parliament • The 'punish a Muslim day' incident • The settlement of Ukrainian refugees in Wales • The Monkeypox outbreak and LGBTQ+ communities • Developing far-right movements in Wales • Events of identity recognition (LGBTQIA+ History Month, Black History Month, International Holocaust Remembrance Day, International Day of Transgender Visibility, International Day of Disabled Persons, Ramadan and International Women's Day) • The Men's and Women's Euros, The World Cup and The Six Nations 2023 The extended pilot period completed in October 2022, and all data services supplied to external clients was paused. The remaining period of the grant (to end of May 2023) is dedicated to the final development stage of the Platform, and the transfer of all technical and data services from the third-party software supplier (Gofore) to SPARK at Cardiff University. |
| Exploitation Route | Several of the pilot client have expressed an interest in paying for continued use of the Platform and services beyond the grant. |
| Sectors | Communities and Social Services/Policy,Digital/Communication/Information Technologies (including Software),Government, Democracy and Justice,Security and Diplomacy |
| Description | In England & Wales police recorded hate crimes are at their highest levels since records began. The migration of hate to the Internet requires the police to address the problem on two fronts. The ESRC funded HateLab (public name for the Cyberhate project) is the first to address the problem both offline and online, generating vital evidence on prevalence, impact and prevention. Lab technologies have been embedded within HMG's NPCC National Online Hate Crime Hub, allowing policymakers and police to prevent hate crime and speech. The HateLab has: v) Innovated by combining social science and computer science research techniques to examine online forms of data to develop an evidence base on online hate speech. Findings revealed that anti-Muslim hate speech spiked in the first 24 hours following terror attacks in 2013 and 2017, and rapidly deescalated, indicating a 'half-life' of 'cyberhate'. In the aftermath of these events social media information flows from police were the second longest lasting within the first 36 hours, indicating that law enforcement online communications might be an effective channel to inform the public, solicit information, and counter rumour, speculation and hate speech; vi) Analysed survey and new forms of data to provide evidence showing hate crime and online hate speech spiked in the final weeks of the Vote Leave and Leave.EU campaigns, following the Brexit vote and at subsequent moments in the Brexit process. These results underpinned the BBC One Panorama documentary 'Hate on the Streets' in 2018; vii) Provided evidence that Brexit related information from Twitter linked to the Russian Internet Research Agency were between 20-40% more likely to be retweeted, compared to UK media, government and public figure/celebrity accounts; viii) Generated evidence that the online abuse of MPs supposedly working against Brexit (so called "mutineers") was organised by a clandestine right-wing group based in London, the results of which appeared on an ITV documentary 'Brexit Online Uncovered' in early 2019. ix) Created an online dashboard that monitors the spread of online hate speech. Using an innovative blend of machine learning (a form of Artificial Intelligence) and social science statistical modelling techniques, the dashboard automatically classifies hateful content in real-time and at a scale hitherto unrealisable. The Online Hate Speech Dashboard was integrated into HMG's National Online Hate Crime Hub (2019). The Hub is the point of contact for all victims of online hate crime, and produces intelligence reports (using the dashboard) for police, senior civil servants and MPs. HateLab results on the spread of online hate speech around events allowed the Hub to better understand the dynamics of propagation, leading to improved response times, better support for victims and more effective allocation of resources. The Director of the Hub states our research has resulted in economic savings of ~£500,000 via the provision of the Dashboard, cloud and data services and an implementation evaluation. HateLab was invited to the Home Affairs Select Committee's inquiry on Hate Crime and its Violent Consequences in November 2016, set-up in response to the murder of Jo Cox MP and the rising levels of hate speech and crimes against the general public and MPs. HateLab evidence was cited in the committee's summary report showing that online hate speech could be detected at scale and in real-time with AI developed at Cardiff. As a result, the inquiry criticised social media companies for not using such technology to counter the spread of hate. HateLab and the Silver Circle law firm, Mischon de Reya, established a partnership in 2018 to publish high-profile reports on the topic of online hate speech, containing legal advice for victims, solicitors and police. The first report was published in early 2019. A co-branded online hate speech 'tracker', available to the public, launched in mid 2019. Evaluations of HateLab Platform Pilot: Welsh Government The Welsh Government's Inclusion and Cohesion Team is responsible for gathering and helping form responses to emerging community and societal tensions at the local, regional and national level in Wales. Its Community Cohesion Programme funds a network of 8 Community Cohesion Coordinators across regions in Wales. As part of their work programme, the Community Cohesion team monitors community tensions in their respective regions, working with the appropriate partners to mitigate these issues as they arise. The Cohesion team provide monthly monitoring reports to Welsh Government but will also send ad hoc updates if the matter is urgent or could potentially escalate. We partnered with HateLab as we have a need to monitor the online social space as part of our Community Cohesion team assessments. Social media, in particular Twitter, is a key space for the discussion of local and regional issues, and without HateLab's technology we had no programmatic way to reliably assess variation in online tensions that relate to offline events. Monitoring exercises using the HateLab Platform: As part of this pilot, we wanted to explore how the HateLab platform could potentially help supplement and enhance the work of the Cohesion teams. Since October 2021, whenever we have been alerted to emerging or potential community tensions that we felt appropriate, we have created queries on the platform to track these developing situations for any flashpoints which could result in hateful communications on social media. Therefore, most of our activity during the pilot has been informed by intelligence supplied by our Cohesion teams. We feel these 'on the ground' sources lend themselves well to the platform, as many of the incidents tended to be localised, with trackable and unique words useful for searching, such as Welsh place names or locations. This has helped to ensure in some cases that we have not collected as much irrelevant information. In some cases, the attached location was one which could be found in multiple countries, sometimes more than one in the same country, such as Newport. With the support from Dr. Arron Cullen, we looked at how to refine our searches. Regarding the types of incidents we have used the platform to follow, it is difficult to provide a case study as the incidents chosen were highly sensitive and it most cases the incidents are ongoing, whether unresolved or still under investigation. In general, we have used the platform to monitor very localised situations/incidents, which in every case, was not a hate incident, but had the potential to attract negative attention and hateful comments online. This approach has meant many of our queries did not identify hateful comments as the situations thankfully did not escalate to that point. This frequent outcome reflected our intended use of the platform. For us, the value of the platform was not just in identifying hate speech, but also to monitor where tensions did not result in hate speech online. An example of this would be the migration of Ukrainians to the UK following the Russian invasion in February 2022. We used HateLab to monitor signs of any negative reaction to the news that Ukrainian families would be relocated to Wales, as there was a chance that far-right groups might capitalise on the situation with anti-migrant rhetoric. Concerns were increased following an incident in Hermon, Pembrokeshire where Ukrainian flags were torn down, as reported by the BBC. While we ran these queries, HateLab did not pick up any examples of hate speech online via the queries we set up. This tallied with the general feeling of goodwill towards Ukrainians that was supported by numerous other sources. Likewise, there was an incident were a far-right group wrongly identified a hotel as a location where asylum seekers would be accommodated. The use of hotels to accommodate asylum seekers is an area which attracts interest from several far-right groups, who use these cases to cause division in local communities. We used HateLab to monitor whether this misinformation would spread and result in localised tensions or hate speech. Again, it did not, which for us was a positive and provided additional reassurance that the situation had passed without escalation. In those cases where situations did escalate, the platform showed us how quickly social media accounts from outside the community/area can take an interest in an incident or developing situation. For example, a protest about a particular cause will likely attract other like-minded accounts via hashtags or influencers. In these cases, we saw far more activity as more users were attracted to the developing situation. As previously mentioned, we are aware that social media users have become more savvy in the way they direct abuse, and often do not use openly hateful language, instead choosing to use more coded words. An example would be the use of 'Ok groomer' towards the LGBTQ+ community and people that speak in favour of relationship and sexuality education in schools. The search function on HateLab provides a way of homing in on these terms through the lens of the anti-LGBTQ+ classifiers, and provides us with a method of widening our searches and generally allowing us to be more dynamic to developing terms or words. Going forward, we have considered how the platform could help us to respond with counter messaging/counter narratives. The word cloud function effectively pulls together common phrases, which in most cases are not overly hateful, but can support or promote stereotypical tropes and 'dog whistle' messaging. For example, our queries relating to asylum seekers have picked up the term 'economic migrants' which pushes a trope around people moving to the UK for financial reasons rather than them seeking sanctuary. With some co-ordination with partners, we would like to explore the use of targeted communications to counter such messaging. Galop (LGBTQIA+ anti-violence charity) For the last 35 years Galop has been working to support the rights of LGBT+ people in the UK, through our work on hate crime, domestic abuse, sexual violence and policing. Our professional casework and helpline services focus on empowerment based support, advice and advocacy to people facing online and offline abuse and violence. We also deliver policy work addressing hate crime by producing research, guidance documents and training. Additionally, we work closely with government, criminal justice, charity and academic bodies to provide advice on issues relating to hate crime. Within the UK we lead the Community Alliance to Combat Hate, an intersectional partnership of leading anti-hate crime charities. Internationally we work closely with the European Commission, OSCE and overseas LGBT+ organisations to stand against online hate speech and hate crime. Monitoring exercises using HateLab platform: The HateLab platform was used by Galop for several months, supported by Dr. Arron Cullen. During this period, Galop ran several different monitoring exercises to track anti-LGBT hate online. These exercises were conducted using common and fewer common hashtags being used by those with anti-LGBT+ views. A sampling exercise was conducted using the platform to gather these hashtags, as well as using common hashtags of abuse from Galop's own service users. Once gathered the monitoring exercises were conducted. The exercises allowed Galop to pick up unknown hashtags that were being used online by those with anti-LGBT+ attitudes, which provided a steppingstone in discovering and tracking a wider set anti-LGBT+ beliefs than was previously possible without the platform. For example, Galop ran a 'Monkeypox' monitoring exercise for several months using the platform. This exercise allowed us to identify when there were spikes in homophobic and transphobic comments being made online and to match these against offline world events. Using the platform, we saw a spike of homophobic comments against gay men when news articles on the outbreak were trending. This fed into Galop's narrative that there was a strong belief online that 'Monkeypox' was a 'gay' illness, fuelled by certain media outlets, that were helping generate homophobic stigma. The HateLab platform proved to be a powerful research tool and Galop was very grateful to be part of the pilot. The learning from the pilot showed that the platform can provide strong evidence on online hate against our communities that is often left undetected. Overall, the HateLab platform is a brilliant resource. National Online Hate Crime Hub The UK National Online Hate Crime Hub was established by the Home Secretary in 2017 to tackle online forms of hate crime, that increased dramatically in the aftermath of the 2016 referendum vote on the future of the UK in the EU. It acts as the point of contact for all victims of online hate crime, and produces intelligence reports. Monitoring exercises using HateLab platform: The provision of the HateLab platform, co-created with the Hub, has fundamentally changed the way we monitor the spread of hate speech during national events. Prior to the platform, Hub staff relied on the Twitter platform interface to gather evidence on the ebb and flow of hate speech around events, such as the referendum vote on the future of the UK in the EU. This proved to be an inadequate method of generating the required insights to track and respond to the problem. During live operations, we were quickly inundated with irrelevant information and failed to capture hate speech in a systematic and reliable way. Through our close collaboration with HateLab, we have co-created technical solutions to overcome these problems. The platform employs sophisticated Machine Learning algorithms to automatically classify hate speech across recognised characteristics at scale and speed, and displays results via a range visualisation tools (frequency chart, top hashtags, topic clusters, geo-location, networks etc). This ensures the Hub can monitor the production and spread of hate speech around events in robust and reliable way. To date, we have used the platform to monitor hate speech around key moments of the Brexit process, including the abuse of MPs, around terror attacks, and most recently around LGBTQ+ pride month. During these events the platform has allowed the staff in the Hub to better understand the dynamics of hate speech propagation, leading to improved response times, better support for victims and more effective allocation of resources. Our latest monitoring exercise around LGBTQ+ pride month in the UK was conducted because our intelligence suggested we would see higher levels of online hate speech during this event than we would typically expect to see throughout the rest of the year. The HateLab platform allowed us to clearly track online hate at the regional and national level during the event. We were able to determine trends on an hourly basis, generate reports centred around keywords and trending hashtags used by those writing the messages, allowing us to create a fuller view of the issue. The platform was an invaluable resource providing a swift analysis of hateful and harmful material in popular discourse allowing us to assess and respond to community threats in real-time. Overall, the platform provides professionals with the best informed assessment of societal tensions and, combined with nearly a decade of observations of the analytics of social media societal tensions, community intelligence and crime trend data, it enables police managers and partners to make the best-informed decisions on deployment and preventative interventions. EE, BT and Saatch & Saatchi EE, BT and Saatchi & Saatchi required data on hateful online messages sent to professional footballers for national anti-hate campaigns (Hope United) during the Women's Euros and the World Cup. Marketing assets using data the HateLab Platform included apparel, TV, OLV, OOH, and Digital. "HateLab were an essential partner for our Hope United campaigns around the Women's Euros and the World Cup. Their knowledge of online harms and their award-winning AI was of huge value and gave us depth of understanding and credible insights we couldn't have otherwise achieved. BT and EE would not hesitate work with HateLab on future projects." BT Group "HateLab were instrumental in providing the data foundation for our Women's Euros campaign for EE Hope United. We were introduced to them through the data driven shirt design project, working with The Mill Experience; whereby players' individual data histories of online abuse impacted the unique designs of their Hope United kits. The team's tracking of hate across the tournament allowed us to confidently talk about the levels of misogynistic hate our Hope United players received during the Euros and reflecting that in reactive press, digital out of home and social which ran over the finals weekend. We are working with the team again, on the World Cup, and the data tracked during that period is a creative jumping off point for a current project - orientated around tackling homophobia within the men's game." Saatchi & Saatchi |
| First Year Of Impact | 2017 |
| Sector | Digital/Communication/Information Technologies (including Software),Government, Democracy and Justice,Security and Diplomacy,Other |
| Impact Types | Societal,Economic,Policy & public services |
| Description | Evidence cited in Home Affairs Select Committee's inquiry on Hate Crime and its Violent Consequences |
| Geographic Reach | National |
| Policy Influence Type | Contribution to a national consultation/review |
| Impact | HateLab project staff were invited to the Home Affairs Select Committee's inquiry on Hate Crime and its Violent Consequences, set-up in response to the murder of Jo Cox MP and the rising levels of hate speech and crimes against the general public and MPs. HateLab evidence was cited in the committee's summary report showing that online hate speech could be detected at scale and in real-time with AI developed with ESRC funding. As a result, the inquiry criticised social media companies for not using such technology to counter the spread of hate. |
| URL | https://publications.parliament.uk/pa/cm201617/cmselect/cmhaff/609/609.pdf |
| Description | Operational impact within National Cyber Hate Crime Hub |
| Geographic Reach | National |
| Policy Influence Type | Influenced training of practitioners or researchers |
| Impact | Impact Letter RE: HateLab Impact on Policing Online Hate in England and Wales Dear Professor Williams, I write this letter to detail the significant impact your HateLab has had on tackling online hate in England and Wales. As a previous senior operational Police Officer and current National Police Chiefs' Council's National Policing Advisor for Hate Crime, I am responsible for coordinating policy and operational activity across all Government departments and criminal justice agencies in their response to hate crime. I also act as the UK Government's 'National Point of Contact' for hate crime to The Organisation for Security and Cooperation in Europe and manage the National Cyber Hate Crime Hub. The Hub was established by the Home Secretary in 2017 to tackle online forms of hate crime, that increased dramatically in the aftermath of the 2016 referendum vote on the future of the UK in the EU. It acts as the point of contact for all victims of online hate crime, and produces intelligence reports for police, senior civil servants and MPs. I sit on the Police and Cross-Government Hate Crime Programme, consisting of all relevant ministers, which coordinates responses to and oversees activity around Government hate crime action plans. While I have seen the benefits from the maintenance of strong relationships between the police and government departments in tackling hate crime, I recognise the need to also engage with academia to tap into the latest evidence and scientific advances in tackling hate. The most recent hate crime action plan 'Action Against Hate: The UK Government's plan for tackling hate crime (2016)' highlights the need for academics and policymakers to come together to tackle the growing problem of hate crime post Brexit. I first became aware of your novel research on online hate at Cardiff University's Hate Crime Symposium in 2012. Since that time we have established a ground-breaking collaboration, involving at the national level, the police, government departments and Cardiff University, to tackle the intractable and growing problem of online hate crime. I have had the pleasure of being a Co-Investigator on several of your ESRC funded research projects, and have experienced the benefit of working closely with your team in co-creating new tools, based on cutting-edge Artificial Intelligence, that generate novel forms of evidence on the nature of the online hate problem. These new technologies, data and insights have transformed the way my team of police officers and civilian staff at the National Cyber Hate Crime Hub monitor and tackle online hate crime. Specifically, the Hub has benefitted from your research on online hate, that uniquely blends computer science with social science methods, and your provision of the Dashboard for monitoring online hate speech in real-time. I have used your research findings on the enablers and inhibiters of online hate speech, the network dynamics of online hate and the effect of counter-narratives on stemming the spread of hate at national and international government and policing events. The analysis you conducted for me on the production and spread of online anti-Muslim hate speech around 'Punish a Muslim Day' transformed my understanding of the problem, and has fed into operational decision making during live police operations. The provision of the HateLab Dashboard, co-created with the Hub, has fundamentally changed the way we monitor the spread of hate speech during national events. Prior to the Dashboard, Hub staff relied on the Twitter platform interface to gather evidence on the ebb and flow of hate speech around events, such as the referendum vote on the future of the UK in the EU. This proved to be an inadequate method of generating the required insights to track and respond to the problem. During live operations, we were quickly inundated with irrelevant information and failed to capture hate speech in a systematic and reliable way. Through our close collaboration with HateLab, we have co-created technical solutions to overcome these problems. The Dashboard employs sophisticated Machine Learning algorithms to automatically classify hate speech across recognised characteristics at scale and speed, and displays results via a range visualisation tools (frequency chart, top hashtags, topic clusters, geo-location, networks etc). This ensures the Hub can monitor the production and spread of hate speech around events in robust and reliable way. To date, we have used the Dashboard to monitor hate speech around key moments of the Brexit process, including the abuse of MPs, and around the Terror attack on London bridge in November 2019. During these events the Dashboard has allowed the staff in the Hub to better understand the dynamics of hate speech propagation, leading to improved response times, better support for victims and more effective allocation of resources. I estimate the Dashboard has saved the police/national government ~£500,000 that would have been spent on a similar solution had our collaboration not materialised. I understand the provision of the HateLab Dashboard is being extended to the public sector (Welsh Government), security sector (UK Counter Terrorism Network) and the civil society sector (Galop, the LGBT+ anti-violence charity) via a significant UKRI Data Service grant. I am certain that each organisation will benefit significantly from their collaboration with the HateLab, and that in the short to long-term, your research and technology will have a positive impact not only on policy and practice, but also on the lives of victims and their families. Yours sincerely, Paul Giannasi, OBE National Policing Advisor for Hate Crime National Police Chiefs' Council |
| Description | Centre for Cyberhate Research & Policy: Real-Time Scalable Methods & Infrastructure for Modelling the Spread of Cyberhate on Social Media |
| Amount | £383,983 (GBP) |
| Funding ID | ES/P010695/1 |
| Organisation | Economic and Social Research Council |
| Sector | Public |
| Country | United Kingdom |
| Start | 04/2017 |
| End | 12/2019 |
| Description | Online Hate Speech Report and Tracker |
| Amount | £4,000 (GBP) |
| Organisation | Mishcon De Reya |
| Sector | Private |
| Country | United Kingdom |
| Start | 09/2018 |
| End | 12/2019 |
| Title | Online Hate Speech Dashboard |
| Description | The Dashboard This tool allows users to access to all open social media feeds (including the Twitter firehose) using a keyword search to identify variation in hate orientated text contained within posts. The dashboard currently allows for the classification of posts containing text that is antagonistic or hateful based on race (anti-black), religion (anti-Muslim), sexual orientation (anti-gay male and female), disability (anti-physical disability), and Jewish identity. Once posts are classified they can be visualised via a suite of tools: a. Real-time and historic modes, allowing end-user to monitor hate speech as it unfolds, and to search back over periods of user data collection for post-hoc analysis b. An interactive hate line chart displaying frequency of tweets, with customisable scale (raw, percentage, log etc). c. An interactive tool for network analysis of hate tweets (where nodes can be selected for further inspection and the production of sub-networks, such as Twitter @mentions, retweets, followers etc.) d. Red/Amber/Green real-time alert system for anomalous spikes in online hate speech above a baseline (defined by user or inferred from average number of hate posts in a given time-frame) e. Tool to identify top N hate hashtags f. Tool to identify top N hate influencers (e.g. top N accounts responsible for N% of hate speech) g. Tool to identify when a top hate user's account is deleted/suspended h. Tool to identify top victim targets (e.g. top N accounts targeted with hate using @mentions) i. Tool to identify Bot accounts with functionality to remove all suspected bots from the analysis and visualization j. Tool to identify links between social media platforms in posts (e.g. frequency of links to far-right open Facebook pages in tweets, far right post on reddit etc.) k. Topic clustering tool, displaying topics detected in posted text and proportion of topics over whole corpus l. Tool to display simple Wordclouds of hate tweets (in addition to topic detection) m. Export tool (sections of dash can be exported) to PDF, image file, bespoke format for end-user n. Demographic estimation of users at an aggregate level (e.g. gender, age) o. Aggregate (e.g. town, city, PFA) geo-location inference plotted on a scalable map (using Lat/Long, user specified location, location name specified in bio etc. - user can specify which are displayed, with all being selectable at once). Individual visualisation tools can be resized and 'toggled' in and out of the view, allowing the user to select the preferred Dashboard set-up for the monitoring task. The suite of tools can also be split over multiple screens to provide the most complete Dashboard set-up. |
| Type Of Material | Improvements to research infrastructure |
| Year Produced | 2019 |
| Provided To Others? | Yes |
| Impact | The Dashboard will be used by HMG's National Online Hate Crime Hub to collect posts from all open social media feeds at set times around predicted and scheduled landmark events, such as the UK's planned exit from the European Union on 29th March 2019. The Purpose The purpose of the Dashboard along with the results and products it produces, is to assist in the identification of 'anomalous' increases in online hate speech in time and space (where geographical information is available) across multiple open social media sources. Results from the Dashboard will be triangulated with other data and intelligence available to the Hub to determine if any increases in online hate speech may be indicative of a rise in community tensions within offline communities or groups. Where offline community tensions can be verified by multiple data sources (including those beyond the Dashboard) the relevant local authorities will be notified. The Data Collection The Dashboard does not permit the identification of individual offending or offenders. Information produced by the Dashboard can only be used for analytical purposes. The outputs of this analysis will be used to inform policy, strategy and decision making with the overall aim of promoting community cohesion. The Dashboard cannot be used to collect evidence for the purpose of criminal proceedings and its use will not to be disclosed and used as evidence. |
| Description | Collaboration with Mumsnet |
| Organisation | Mumsnet |
| Country | United Kingdom |
| Sector | Charity/Non Profit |
| PI Contribution | We are developing new hate speech classifier to assist Mumsnet with their moderation efforts. |
| Collaborator Contribution | Mumsnet are providing us with over 20 years of deleted posts to train the new hate speech classifiers. |
| Impact | Machine learning classifiers for hate speech. Their creation was a multi-disciplinary effort. |
| Start Year | 2021 |
| Description | Collaboration with Tell MAMA |
| Organisation | Faith Matters |
| Country | United Kingdom |
| Sector | Charity/Non Profit |
| PI Contribution | We are developing new anti-Muslim hate classifiers for Tell MAMA within Faith Matters |
| Collaborator Contribution | Tell MAMA are providing us with data from report of online hate crime |
| Impact | Machine Learning classifiers to detect anti-Muslim hate speech |
| Start Year | 2021 |
| Title | Online Hate Speech Dashboard |
| Description | The Dashboard This tool allows users to access to all open social media feeds (including the Twitter firehose) using a keyword search to identify variation in hate orientated text contained within posts. The dashboard currently allows for the classification of posts containing text that is antagonistic or hateful based on race (anti-black), religion (anti-Muslim), sexual orientation (anti-gay male and female), disability (anti-physical disability), and Jewish identity. Once posts are classified they can be visualised via a suite of tools: a. Real-time and historic modes, allowing end-user to monitor hate speech as it unfolds, and to search back over periods of user data collection for post-hoc analysis b. An interactive hate line chart displaying frequency of tweets, with customisable scale (raw, percentage, log etc). c. An interactive tool for network analysis of hate tweets (where nodes can be selected for further inspection and the production of sub-networks, such as Twitter @mentions, retweets, followers etc.) d. Red/Amber/Green real-time alert system for anomalous spikes in online hate speech above a baseline (defined by user or inferred from average number of hate posts in a given time-frame) e. Tool to identify top N hate hashtags f. Tool to identify top N hate influencers (e.g. top N accounts responsible for N% of hate speech) g. Tool to identify when a top hate user's account is deleted/suspended h. Tool to identify top victim targets (e.g. top N accounts targeted with hate using @mentions) i. Tool to identify Bot accounts with functionality to remove all suspected bots from the analysis and visualization j. Tool to identify links between social media platforms in posts (e.g. frequency of links to far-right open Facebook pages in tweets, far right post on reddit etc.) k. Topic clustering tool, displaying topics detected in posted text and proportion of topics over whole corpus l. Tool to display simple Wordclouds of hate tweets (in addition to topic detection) m. Export tool (sections of dash can be exported) to PDF, image file, bespoke format for end-user n. Demographic estimation of users at an aggregate level (e.g. gender, age) o. Aggregate (e.g. town, city, PFA) geo-location inference plotted on a scalable map (using Lat/Long, user specified location, location name specified in bio etc. - user can specify which are displayed, with all being selectable at once). Individual visualisation tools can be resized and 'toggled' in and out of the view, allowing the user to select the preferred Dashboard set-up for the monitoring task. The suite of tools can also be split over multiple screens to provide the most complete Dashboard set-up. |
| Type Of Technology | Webtool/Application |
| Year Produced | 2019 |
| Impact | The Dashboard will be used by HMG's National Online Hate Crime Hub to collect posts from all open social media feeds at set times around predicted and scheduled landmark events, such as the UK's planned exit from the European Union on 29th March 2019. The Purpose The purpose of the Dashboard along with the results and products it produces, is to assist in the identification of 'anomalous' increases in online hate speech in time and space (where geographical information is available) across multiple open social media sources. Results from the Dashboard will be triangulated with other data and intelligence available to the Hub to determine if any increases in online hate speech may be indicative of a rise in community tensions within offline communities or groups. Where offline community tensions can be verified by multiple data sources (including those beyond the Dashboard) the relevant local authorities will be notified. The Data Collection The Dashboard does not permit the identification of individual offending or offenders. Information produced by the Dashboard can only be used for analytical purposes. The outputs of this analysis will be used to inform policy, strategy and decision making with the overall aim of promoting community cohesion. The Dashboard cannot be used to collect evidence for the purpose of criminal proceedings and its use will not to be disclosed and used as evidence. |
| Description | BBC One Panorama 'Hate on the Streets' |
| Form Of Engagement Activity | A broadcast e.g. TV/radio/film/podcast (other than news/press) |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Media (as a channel to the public) |
| Results and Impact | We participated in BBC One's Panorama 'Hate on the Streets'. The project supplied key evidence on the trends in offline hate crimes following the Brexit vote. The documentary was watched by over 3.4 million. |
| Year(s) Of Engagement Activity | 2018 |
| URL | https://www.youtube.com/watch?v=w-jOhDbrQjQ&feature=youtu.be |
| Description | ITV Exposure 'Brexit Online Uncovered' |
| Form Of Engagement Activity | A broadcast e.g. TV/radio/film/podcast (other than news/press) |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Media (as a channel to the public) |
| Results and Impact | We provided key evidence to ITV's Exposure documentary 'Brexit Online Uncovered' that showed the links between Twitter users who were abusing MPs online, and how press headlines were statistically associated with increases in general online hate speech related to Brexit. The documentary had its premiere in the Houses of Parliament hosted by the Rt Hon Antoinette Sandbach MP for Eddisbury. It was viewed by over 4.1 million. |
| Year(s) Of Engagement Activity | 2019 |
| URL | https://www.youtube.com/watch?v=BcMYxP9zfVU&feature=youtu.be |
| Description | ITV NEWS Special Report on the rise of online hate speech |
| Form Of Engagement Activity | A broadcast e.g. TV/radio/film/podcast (other than news/press) |
| Part Of Official Scheme? | No |
| Geographic Reach | National |
| Primary Audience | Public/other audiences |
| Results and Impact | We provided key evidence to an ITV NEWS special report on the rise of online hate speech. It was broadcast nationally on the lunchtime and evening ITV NEWS shows in early March 2020. Estimated audience over both shows ~6 million. |
| Year(s) Of Engagement Activity | 2020 |
| URL | https://www.youtube.com/watch?v=sBHclogub6M&feature=youtu.be |
| Description | Paper Presented at The Web Conference 2019, San Francisco, CA, USA |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Public/other audiences |
| Results and Impact | Liu, H.et al. 2019. Fuzzy multi-task learning for hate speech type identification. Presented at: The Web Conference 2019, San Francisco, CA, USA, 13-17 May 2019Proceedings of the 2019 World Wide Web Conference. ACM, (10.1145/3308558.3313546) |
| Year(s) Of Engagement Activity | 2019 |
| Description | Paper presented at 1st International Conference on Cyber Deviance Detection (CyberDD) in conjunction with 10th ACM International Conference on Web Search and Data Mining (WSDM 2017), Cambridge, UK |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Burnap, P. and Williams, M. L. 2017. Classifying and modeling cyber hate speech: research and opportunities for practical intervention. Presented at: 1st International Conference on Cyber Deviance Detection (CyberDD) in conjunction with 10th ACM International Conference on Web Search and Data Mining (WSDM 2017), Cambridge, UK, 10 Feb 2017. |
| Year(s) Of Engagement Activity | 2017 |
| Description | Paper presented at Cambridge Institute of Criminology Seminar Series, University of Cambridge |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | National |
| Primary Audience | Public/other audiences |
| Results and Impact | Williams, M. L. and Burnap, P. 2017. Social data science & criminology: machine classification and modelling of cyberhate in online social networks. Presented at: Cambridge Institute of Criminology Seminar Series, University of Cambridge, UK, 9 February 2017. |
| Year(s) Of Engagement Activity | 2017 |
| Description | Paper presented at Data Science and Government Conference, Oxford, UK |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Burnap, P. and Williams, M. L. 2016. Computational human and cyber security analytics for government and policy. Presented at: Data Science and Government Conference, Oxford, UK, 22 June 2016. |
| Year(s) Of Engagement Activity | 2016 |
| Description | Paper presented at Home Office Crime and Policing Analysis Unit Seminar Series, Westminster, London, UK |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | National |
| Primary Audience | Policymakers/politicians |
| Results and Impact | Williams, M. L. and Burnap, P. 2017. Detecting crime events using social media. Presented at: Home Office Crime and Policing Analysis Unit Seminar Series, Westminster, London, UK, July, 2017. |
| Year(s) Of Engagement Activity | 2017 |
| Description | Paper presented at International Conference on Machine Learning and Cybernetics, Chengdu, China |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Public/other audiences |
| Results and Impact | Alorainy, W.et al. 2018. Suspended accounts: A source of Tweets with disgust and anger emotions for augmenting hate speech data sample. Presented at: International Conference on Machine Learning and Cybernetics, Chengdu, China, 15-18 July 2018. |
| Year(s) Of Engagement Activity | 2018 |
| Description | Paper presented at Internet Leadership Academy, Oxford Internet Institute, University of Oxford |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Williams, M. and Burnap, P. 2017. Online extremism and hate speech: definition, measurement & regulation. Presented at: Internet Leadership Academy, Oxford Internet Institute, University of Oxford, UK, 26 September 2017. |
| Year(s) Of Engagement Activity | 2017 |
| Description | Paper presented at Jensen Lecture Series, Duke University, NC, US |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Public/other audiences |
| Results and Impact | Williams, M. L. 2016. Crime sensing with big data: the affordances and limitations of using open source communications to estimate crime patterns. Presented at: Jensen Lecture Series, Duke University, NC, US, 2016. |
| Year(s) Of Engagement Activity | 2016 |
| Description | Paper presented at UK Government Data Science Community Interest Workshop, ONS Data Science Campus, Newport, Wales, UK |
| Form Of Engagement Activity | A talk or presentation |
| Part Of Official Scheme? | No |
| Geographic Reach | National |
| Primary Audience | Professional Practitioners |
| Results and Impact | Williams, M. L. and Burnap, P. 2017. Data science solutions for detecting and monitoring Brexit related online hate speech. Presented at: UK Government Data Science Community Interest Workshop, ONS Data Science Campus, Newport, Wales, UK, 4 September 2017. |
| Year(s) Of Engagement Activity | 2017 |
| Description | Paper presented at: SERENE-RISC Workshop, Université de Montréal, Montreal, QC, Canada |
| Form Of Engagement Activity | Participation in an activity, workshop or similar |
| Part Of Official Scheme? | No |
| Geographic Reach | International |
| Primary Audience | Professional Practitioners |
| Results and Impact | Williams, M. L. 2017. Big Data and criminology: Research from the UK. Presented at: SERENE-RISC Workshop, Université de Montréal, Montreal, QC, Canada, 26 April 2017. |
| Year(s) Of Engagement Activity | 2017 |