Constrained conditional model

Abstract: : In this thesis, we study the importance of background knowledge in relation extraction systems We not only demonstrate the bene ts of leveraging background knowledge to improve the systems' performance but also propose a principled framework that allows one to effectively incorporate knowledge into statistical machine learning models for relation extraction Our work is motivated by the fact that relation extraction systems in the literature usually use evidence that is written explicitly in the input text to detect and characterize the semantic relations between target concepts Although this approach achieves reasonable performance, it does not necessarily guarantee accurate extraction due to problems of poor information representation of the systems' inputs and lack of knowledge to support logical reasoning We argue that relation extraction systems would benefit from using one or more background knowledge sources, both in enriching the systems' inputs and biasing the final outputs We illustrate our framework in the context of several learning-based relation extraction tasks The first task is Taxonomic Relation identification where we employ an external knowledge source to construct meaning representation of the task inputs and support global inference to identify taxonomic relations between input terms In the second task, Event Relation Discovery, we focus on identify causality relation between events in text Our approach leverages background knowledge to perform joint inference among several classifiers that make local decisions on event causality relation After that, we study the problem of constructing a timeliine of events extracted from text, Event Timeline Construction To address this task, we propose a new timeline representation with events mapped to absolute time intervals In this work, we present a time interval-based global inference model that jointly assigns events into time intervals on a timeline and orders events temporally

Abstract: This paper investigates one of the most fundamental computer vision problems: image segmentation. We propose a supervised hierarchical approach to object-independent image segmentation. Starting with over-segmenting superpixels, we use a tree structure to represent the hierarchy of region merging, by which we reduce the problem of segmenting image regions to finding a set of label assignment to tree nodes. We formulate the tree structure as a constrained conditional model to associate region merging with likelihoods predicted using an ensemble boundary classifier. Final segmentations can then be inferred by finding globally optimal solutions to the model efficiently. We also present an iterative training and testing algorithm that generates various tree structures and combines them to emphasize accurate boundaries by segmentation accumulation. Experiment results and comparisons with other very recent methods on six public data sets demonstrate that our approach achieves the state-of-the-art region accuracy and is very competitive in image segmentation without semantic priors.