New event detection, also known as first story detection (FSD), has become very popular in recent years. The task consists of finding previously unseen events from a stream of documents. Despite the apparent simplicity, FSD is very challenging and has applications anywhere where timely access to fresh information is crucial: from journalism to stock market trading, homeland security, or emergency response. With the rise of user generated content and citizen journalism we have entered an era of big and noisy data, yet traditional approaches for solving FSD are not designed to deal with this new type of data. The amount of information that is being generated today exceeds by many orders of magnitude previously available datasets, making traditional approaches obsolete for modern event detection. In this thesis, we propose a modern approach to event detection that scales to unbounded streams of text, without sacrificing accuracy. This is a crucial property that enables us to detect events from large streams like Twitter, which none of the previous approaches were able to do. One of the major problems in detecting new events is vocabulary mismatch, also known as lexical variation. This problem is characterized by different authors using different words to describe the same event, and it is inherent to human language. We show how to mitigate this problem in FSD by using paraphrases. Our approach that uses paraphrases achieves state-of-the-art results on the FSD task, while still maintaining efficiency and being able to process unbounded streams. Another important property of user generated content is the high level of noise, and Twitter is no exception. This is another problem that traditional approaches were not designed to deal with, and here we investigate different methods of reducing the amount of noise. We show that by using information from Wikipedia, it is possible to significantly reduce the amount of spurious events detected in Twitter, while maintaining a very small latency in detection. A question is often raised as to whether Twitter is at all useful, especially if one has access to a high-quality stream such as the newswire, or if it should be considered as sort of a poor man’s newswire. In our comparison of these two streams we find that Twitter contains events not present in the newswire, and that it also breaks some events sooner, showing that it is useful for event detection, even in the presence of newswire.

In today’s world the internet and social media are omnipresent and information is accessible to everyone. This shifted the advantage from those who have access to information to those who do so first. Identifying new events as they emerge is of substantial value to financial institutions who consider realtime information in their decision making processes, as well as for journalists that report about breaking news and governmental agencies that collect information and respond to emergencies. First Story Detection is the task of identifying those documents in a stream of documents that talk about new events first. This seemingly simple task is non-trivial as the computational effort increases with every processed document. Standard approaches to solve First Story Detection determine a document’s novelty by comparing it to previously seen documents. This results in the highest reported accuracy but even the currently fastest system only scales to 10% of the Twitter stream. In this thesis, we propose a new algorithm family, called memory-based methods, able to scale to the full Twitter stream on a single core. Our memory-based method computes a document’s novelty up to two orders of magnitude faster than state-of-the-art systems without sacrificing accuracy. This thesis additional provides original work on the impact of processing unbounded data streams on detection accuracy. Our experiments reveal for the first time that the novelty scores of state-of-the-art comparison based and memory-based methods decay over time. We show how to counteract the discovered novelty decay and increase detection accuracy. Additionally, we show that memory-based methods are applicable beyond First Story Detection by building the first real time rumour detection system on social media streams.

abstract: A major challenge in automated text analysis is that different words are used for related concepts. Analyzing text at the surface level would treat related concepts (i.e. actors, actions, targets, and victims) as different objects, potentially missing common narrative patterns. Generalized concepts are used to overcome this problem. Generalization may result into word sense disambiguation failing to find similarity. This is addressed by taking into account contextual synonyms. Concept discovery based on contextual synonyms reveal information about the semantic roles of the words leading to concepts. Merger engine generalize the concepts so that it can be used as features in learning algorithms. / Dissertation/Thesis / Masters Thesis Computer Science 2015