Periodic Reporting for period 2 - COMRADES (Collective Platform for Community Resilience and Social Innovation during Crises)
Reporting period: 2017-07-01 to 2018-12-31
Response to crisis often reveals organisational and technological shortcomings, which threaten community recovery and sustainable. Even though some technological solutions exist, challenges of communication, interoperability, and data analytics remain. The deployment and use of technologies, and the social structures in which they are adopted, are interdependent. Hence it is imperative to develop human-centred technologies that take into account actual real world practices of affected populations and responders. COMRADE’s main objectives are:
- Extract the socio-technical requirements for collective resilience platforms
- Automatically identify, process, assess, and monitor emergency events in distributed communities and social media sources
- Measure the informativeness and validity of crisis information
- Develop and deploy the intelligent COMRADES platform for community resilience
- Requirements gathering and evaluation: We thoroughly examined literature on resilience, and the role of communities and information, and gathered extensive requirements from communities through surveys, interviews, focus groups, and workshops, and launched several evaluation exercises of the COMRADES platform and tools.
- Veracity and informativeness: COMRADES researched and developed various analysis tools to process textual reports, and to measure their veracity and informativeness. We developed and released a social media version of YODIE; a tool for multilingual information extraction that is designed for named entity recognition, disambiguation and linking. Also developed a tool for classifying informativeness and actionability, named EMINA (Emergent Informativeness and Actionability), and another tool for measuring veracity. Both tools are integrated into the COMRADES platform.
- Crises data analysis: COMRADES developed tools for classifying crisis-related information based on relevancy and event types. We constructed a three-tier ontology DoRES (Document-Report-Event-Situation) for representing information on crisis-related events, and researched, developed, and integrated a Crisis Event Extraction Service (CREES); a semantic deep-learning method for automatically detecting crisis-related events in social media data. We also experimented with novel designs of models for automatically classifying crisis information that can be trained on data about certain types of crises, and languages, and applied to other crises and languages. Furthermore, we also experiment with different approaches for matching crisis event content.
- COMRADES platform: The platform forms an extension of the Ushahidi platform, by integrating it with the tools above. The platform is available on Github, and fully integrates YODIE, EMINA, and Veracity services from, and CREES, and meets the main requirements gathered in the project. Platform evaluation was performed and documented in several deliverables.
- Dissemination, exploitation, and communication: Overall the project published more than 30 scientific articles in peer reviewed conferences and journals. And launched several workshops, newsletters, and public engagement events.
The main exploitable results of COMRADES are:
- CREES - An advanced Artificial Intelligent (A.I.) method that identifies crisis events in social media streams which is called Crisis Event Extraction Service (CREES). CREES is available as an integrated component in the COMRADES platform, but also as an API and as a Google Sheet plugin (https://tinyurl.com/y67523rf).
- Informativeness classifier for social media posts - The classifier filters the deluge of data on social media and selects which messages are more informative and which are more actionable (https://github.com/GateNLP/emina).
- Entity finder for social media posts - The tool automatically identify entities in given posts, to help digital responders and others to very quickly learn about the locations, organisations, and other entities mentioned in the post. https://gate.ac.uk/applications/yodie.html
- Veracity classifier for social media messages - Automatic method for calculating the probable veracity level of a given message. This tool is designed to help users with assessing the veracity of messages posted in social media, such as in Twitter. It is deployed as a service (https://cloud.gate.ac.uk/shopfront/displayItem/rumour-veracity).
- Ushahidi platform that allows organizations to gather information, incidents, and reports from people on the ground, gain situational awareness, and then respond more efficiently and effectively. A “deployment” in Ushahidi’s context is a hosted instance of the Ushahidi Platform. Ushahidi has created a COMRADES deployment for use by the consortium partners that is accessible at: https://comrades.ushahidi.com.
- Resilience: project explored the integration of resilience theories and conceptual frameworks with collective platforms for community empowerment and recovery from a variety of crises situations. With a community-driven design approach, new socio-technical requirements emerged.
- Content and Source Validity Assessment: COMRADES developed novel online machine learning methods to detect low credibility information, in multiple languages.
- Entity Extraction: the ambiguity of entities mentioned in social media posts are resolved using LOD-based entity disambiguation tools and the accumulation of context provided by incoming additional tweets relevant to the crisis. Dealing with a streaming scenario, where new information becomes available over time, as opposed to most previous work which involves a post-hoc batch analysis system, means that the state of knowledge about a crisis is fluid and assertions may become better-supported or contradicted over time.
- Crisis Event Modelling detection: COMRADES constructed semantic models for representing, integrating and reasoning over the various types and subjects of micro events that emerge during crisis situations. The project provided models for emergency event identification and matchmaking, which apples advance semantic similarity algorithms to determine the strength of event pairings. Crisis information classification methods were designed to function across different type of crises and on data in different languages.