Ao Feng

Event threading within news topics

With the overwhelming volume of online news available today, there is an increasing need for automatic techniques to analyze and present news to the user in a meaningful and efficient manner. Previous research focused only on organizing news stories by their topics into a flat hierarchy. We believe viewing a news topic as a flat collection of stories is too restrictive and inefficient for a user to understand the topic quickly. In this work, we attempt to capture the rich structure of events and their dependencies in a news topic through our event models. We call the process of recognizing events and their dependencies <i>event threading</i>. We believe our perspective of modeling the structure of a topic is more effective in capturing its semantics than a flat list of on-topic stories. We formally define the novel problem, suggest evaluation metrics and present a few techniques for solving the problem. Besides the standard word based features, our approaches take into account novel features such as temporal locality of stories for event recognition and time-ordering for capturing dependencies. Our experiments on a manually labeled data sets show that our models effectively identify the events and capture dependencies among them.

Finding and linking incidents in news

News reports are being produced and disseminated in overwhelming volume, making it difficult to keep up with the newest information. Most previous research in automatic news organization treated news topics as a flat list, ignoring the intrinsic connection among individual reports. We argue that more contextual information within and across the topics will benefit users in their news understanding process. A news organization infrastructure, incident threading, is proposed in this article. All text snippets describing the occurrence of a real-world happening are combined into a news incident, and a network is composed of incidents that are interconnected by links in certain types. A limited vocabulary of connection types is defined and corresponding rules are established based upon the human experience of news understanding. The incident threading system is implemented with two different algorithms. One starts from clustering of text passages and then creates links with pre-built rules. The other method defines a global score function over the whole collection and solves the optimization problem with simulated annealing. The former achieves higher accuracy in the identification of incidents and the latter generates better links, which is preferred since the links are more important for the formation of the incident network.

A critical examination of TDT's cost function

Topic Detection and Tracking (TDT) tasks are evaluated using a cost function. The standard TDT cost function assumes a constant probability of relevance P(rel) across all topics. In practice, P(rel) varies widely across topics. We argue using both theoretical and experimental evidence that the cost function should be modified to account for the varying P(rel).

Incident threading for news passages

With an overwhelming volume of news reports currently available, there is an increasing need for automatic techniques to analyze and present news to a general reader in a meaningful and efficient manner. We explore incident threading as a possible solution to this problem. All text that describes the occurrence of a real-world happening is merged into a news incident, and incidents are organized in a network with dependencies of predefined types. Earlier attempts at this problem have assumed that a news story covers a single topic. We move beyond that limitation to introduce passage threading, which processes news at the passage level. First we develop a new testbed for this research and extend the evaluation methods to address new granularity issues. Then a three-stage algorithm is described that identifies on-subject passages, groups them into incidents, and establishes links between related incidents. Finally, we observe significant improvement over earlier work when we optimize the harmonic mean of the appropriate evaluation measures. The resulting performance exceeds the level that a calibration study shows is necessary to support a reading comprehension task.

Global denoising for 3D MRI

Denoising is the primary preprocessing step before subsequent clinical diagnostic analysis of MRI data. Common patch-based denoising methods rely heavily on the degree of patch matching, which limits their performance by the necessity of finding sufficiently similar patches. In this paper, we propose a global filtering framework, in which each voxel is restored with information from the whole 3D image. This global filter is not restricted to any specific patchbased filter, as it is a low-rank approximation using the Nyström method combined with a low sampling rate and a kmeans clustering adaptive sampling scheme. Experiments demonstrate that this method utilizes information effectively from the whole image for denoising, and the framework can be applied on top of most patch-based methods to further improve the performance.