• Distributed event detection based on state-of-the-art locality hashing technologies, enabling scalable processing of high-volume social media streams on cheaper commodity hardware.
• Extended using word-embedding techniques, reducing detection error rates by around 9% absolute over state-of-the-art.

• Uses n-gram graphs to tackle vocabulary mismatch between the training instances and social media posts. N-gram graphs sentiment analysis method has accuracy 85% and it is more effective than current state-of-the-art approaches that have approximately an accuracy of 80%.
• Novel approach for targeted sentiment classification, which enables users to track sentiments regarding people or organisations more accurately

• A new algorithm was developed to re-construct discussion threads/reply chains from Twitter.
• Development of a `reply' index, which maintains the chain of reply relationships for each post in an efficient manner.
• Discussion Thread Exploration User Interface

• An automatic rumour post classifier was developed which achieved between ~0.65 F1 and ~0.75 F1 effectiveness (over three test events).
• A novel reverse image search system was developed to automatically identify images posted that belong to past events.
• Crowdsourcing was examined as a human-annotated means for identifying rumours on-demand, showing that crowdsourced assessors are able to quickly identify rumour-related tweets from their text with high agreement.
• A state-of-the-art automatic credibility estimation approach was extended to consider hyperlinked evidence, which resulted in an increase in classification accuacy by 4-5%.
• A range of rumour and credibility identification datasets were developed that can be used by future researchers.

• The Virtual Spaces Management Layer (VSML) which includes a set of adaptors for Facebook and Twitter and a central component for aggregating data from them.
• The Experiment Management Layer (EML) provides tools for managing experiments, storing experiments data and it distributes them to the data analysis service.
• The Data Analysis service illustrates the results from experiments in visual form and helps policy makers to interpret the results. The presentation layer of the data analysis service provides interactive elements, which allows filtering of results. Data analysis of running experiments happens in near real-time.
• The +Spaces Front-End is a graphical interface for End Users.

• Topical CI based on topics that are discussed by social network users according to their published textual posts.
• Topological CI based on interactions, such as reply, mention, retweet and follow, between users.

• Simultaneously monitor a great number of events across social media channels
• Permanent storage of events by the platform. This a major advantage in the case an event has been disabled or deleted by the initiator
• Advanced filters and facets to search only for relevant and actual content.

• A novel real-time summarization system named xQuADSumm was developed that leverages explicit diversification techniques to produce updating summaries based on either manually or automatically generated event type facets.
• xQuADSumm outperforms the commonly used ‘rank-then-select’ baseline approach by 21% in terms of effectiveness.
• xQuADSumm is efficient and scalable, as it is able to process between 2200 and 2500 posts per second per machine allocated.

• The Integrated Search component was developed on-top of the Terrier IR platform
• This provides new and more effective document weighting models designed for use on short texts
• A new distributed stream processing version of the platform was developed that can scale the indexing and retrieval over multiple commodity machines using Apache Storm

• Alternative communication channel for coordination of in-field operations
• Integrated with COP software
• Ability to reach high amount of individuals with low effort at COP software

According to whether the messages contain links, mentions to other users or if they are retweet (in case of messages crawled from Twitter), users can be classified as follows:

• Endogenous: users post mostly original tweets with rough text (without links or mentions to other users)
• Conversationals: users post mostly original tweets containing mentions to other users, and no links.
• Generalists: users post different kinds of tweets, without a marked preference for any class.
• Echoers: users post mostly retweets of other people’s messages which do not contain links.
• Link feeders: users post especially tweets containing links (URLs).

• Uses various data fusion techniques that satisfy in a better way the social sensors during a crisis event.
• Reasoning techniques are tested to justify the rumour spreading and sentiment dispositions of the communities.