Probabilistic Models for Scalable Knowledge Graph Construction

In the past decade, systems that extract information from millions of Internet documents have become commonplace. Knowledge graphs -- structured knowledge bases that describe entities, their attributes and the relationships between them -- are a powerful tool for understanding and organizing this vast amount of information. However, a significant obstacle to knowledge graph construction is the unreliability of the extracted information, due to noise and ambiguity in the underlying data or errors made by the extraction system and the complexity of reasoning about the dependencies between these noisy extractions. My dissertation addresses these challenges by exploiting the interdependencies between facts to improve the quality of the knowledge graph in a scalable framework. I introduce a new approach called knowledge graph identification (KGI), which resolves the entities, attributes and relationships in the knowledge graph by incorporating uncertain extractions from multiple sources, entity co-references, and ontological constraints. I define a probability distribution over possible knowledge graphs and infer the most probable knowledge graph using a combination of probabilistic and logical reasoning. Such probabilistic models are frequently dismissed due to scalability concerns, but my implementation of KGI maintains tractable performance on large problems through the use of hinge-loss Markov random fields, which have a convex inference objective. This allows the inference of large knowledge graphs using 4M facts and 20M ground constraints in 2 hours. To further scale the solution, I develop a distributed approach to the KGI problem which runs in parallel across multiple machines, reducing inference time by 90%. Finally, I extend my model to the streaming setting, where a knowledge graph is continuously updated by incorporating newly extracted facts. I devise a general approach for approximately updating inference in convex probabilistic models, and quantify the approximation error by defining and bounding inference regret for online models. Together, my work retains the attractive features of probabilistic models while providing the scalability necessary for large-scale knowledge graph construction. These models have been applied on a number of real-world knowledge graph projects, including the NELL project at Carnegie Mellon and the Google Knowledge Graph.

Noisy Monte Carlo: convergence of Markov chains with approximate transition kernels

Monte Carlo algorithms often aim to draw from a distribution π by simulating a Markov chain with transition kernel P such that π is invariant under P. However, there are many situations for which it is impractical or impossible to draw from the transition kernel P. For instance, this is the case with massive datasets, where is it prohibitively expensive to calculate the likelihood and is also the case for intractable likelihood models arising from, for example, Gibbs random fields, such as those found in spatial statistics and network analysis. A natural approach in these cases is to replace P by an approximation Pˆ. Using theory from the stability of Markov chains we explore a variety of situations where it is possible to quantify how ’close’ the chain given by the transition kernel Pˆ is to the chain given by P . We apply these results to several examples from spatial statistics and network analysis.

A novel framework of training hidden Markov support vector machines from lightly-annotated data

Natural language understanding (NLU) aims to map sentences to their semantic mean representations. Statistical approaches to NLU normally require fully-annotated training data where each sentence is paired with its word-level semantic annotations. In this paper, we propose a novel learning framework which trains the Hidden Markov Support Vector Machines (HM-SVMs) without the use of expensive fully-annotated data. In particular, our learning approach takes as input a training set of sentences labeled with abstract semantic annotations encoding underlying embedded structural relations and automatically induces derivation rules that map sentences to their semantic meaning representations. The proposed approach has been tested on the DARPA Communicator Data and achieved 93.18% in F-measure, which outperforms the previously proposed approaches of training the hidden vector state model or conditional random fields from unaligned data, with a relative error reduction rate of 43.3% and 10.6% being achieved.

Bayesian hierarchical models for analysing spatial point-based data at a grid level: A comparison of approaches

Spatial data are now prevalent in a wide range of fields including environmental and health science. This has led to the development of a range of approaches for analysing patterns in these data. In this paper, we compare several Bayesian hierarchical models for analysing point-based data based on the discretization of the study region, resulting in grid-based spatial data. The approaches considered include two parametric models and a semiparametric model. We highlight the methodology and computation for each approach. Two simulation studies are undertaken to compare the performance of these models for various structures of simulated point-based data which resemble environmental data. A case study of a real dataset is also conducted to demonstrate a practical application of the modelling approaches. Goodness-of-fit statistics are computed to compare estimates of the intensity functions. The deviance information criterion is also considered as an alternative model evaluation criterion. The results suggest that the adaptive Gaussian Markov random field model performs well for highly sparse point-based data where there are large variations or clustering across the space; whereas the discretized log Gaussian Cox process produces good fit in dense and clustered point-based data. One should generally consider the nature and structure of the point-based data in order to choose the appropriate method in modelling a discretized spatial point-based data.

An external field prior for the hidden Potts model with application to cone-beam computed tomography

In images with low contrast-to-noise ratio (CNR), the information gain from the observed pixel values can be insufficient to distinguish foreground objects. A Bayesian approach to this problem is to incorporate prior information about the objects into a statistical model. A method for representing spatial prior information as an external field in a hidden Potts model is introduced. This prior distribution over the latent pixel labels is a mixture of Gaussian fields, centred on the positions of the objects at a previous point in time. It is particularly applicable in longitudinal imaging studies, where the manual segmentation of one image can be used as a prior for automatic segmentation of subsequent images. The method is demonstrated by application to cone-beam computed tomography (CT), an imaging modality that exhibits distortions in pixel values due to X-ray scatter. The external field prior results in a substantial improvement in segmentation accuracy, reducing the mean pixel misclassification rate for an electron density phantom from 87% to 6%. The method is also applied to radiotherapy patient data, demonstrating how to derive the external field prior in a clinical context.

Modelling of solute transport in a mild heterogeneous porous medium using stochastic finite element method: Effects of random source conditions

Randomness in the source condition other than the heterogeneity in the system parameters can also be a major source of uncertainty in the concentration field. Hence, a more general form of the problem formulation is necessary to consider randomness in both source condition and system parameters. When the source varies with time, the unsteady problem, can be solved using the unit response function. In the case of random system parameters, the response function becomes a random function and depends on the randomness in the system parameters. In the present study, the source is modelled as a random discrete process with either a fixed interval or a random interval (the Poisson process). In this study, an attempt is made to assess the relative effects of various types of source uncertainties on the probabilistic behaviour of the concentration in a porous medium while the system parameters are also modelled as random fields. Analytical expressions of mean and covariance of concentration due to random discrete source are derived in terms of mean and covariance of unit response function. The probabilistic behaviour of the random response function is obtained by using a perturbation-based stochastic finite element method (SFEM), which performs well for mild heterogeneity. The proposed method is applied for analysing both the 1-D as well as the 3-D solute transport problems. The results obtained with SFEM are compared with the Monte Carlo simulation for 1-D problems.

Doubly spectral stochastic finite element method (DSSFEM) for random field problems

Uncertainties in complex dynamic systems play an important role in the prediction of a dynamic response in the mid- and high-frequency ranges. For distributed parameter systems, parametric uncertainties can be represented by random fields leading to stochastic partial differential equations. Over the past two decades, the spectral stochastic finite-element method has been developed to discretize the random fields and solve such problems. On the other hand, for deterministic distributed parameter linear dynamic systems, the spectral finite-element method has been developed to efficiently solve the problem in the frequency domain. In spite of the fact that both approaches use spectral decomposition (one for the random fields and the other for the dynamic displacement fields), very little overlap between them has been reported in literature. In this paper, these two spectral techniques are unified with the aim that the unified approach would outperform any of the spectral methods considered on their own. An exponential autocorrelation function for the random fields, a frequency-dependent stochastic element stiffness, and mass matrices are derived for the axial and bending vibration of rods. Closed-form exact expressions are derived by using the Karhunen-Loève expansion. Numerical examples are given to illustrate the unified spectral approach.

Estudo da quantificação de incertezas para o problema de contaminação de meios porosos heterogêneos

As técnicas de injeção de traçadores têm sido amplamente utilizadas na investigação de escoamentos em meios porosos, principalmente em problemas envolvendo a simulação numérica de escoamentos miscíveis em reservatórios de petróleo e o transporte de contaminantes em aquíferos. Reservatórios subterrâneos são em geral heterogêneos e podem apresentar variações significativas das suas propriedades em várias escalas de comprimento. Estas variações espaciais são incorporadas às equações que governam o escoamento no interior do meio poroso por meio de campos aleatórios. Estes campos podem prover uma descrição das heterogeneidades da formação subterrânea nos casos onde o conhecimento geológico não fornece o detalhamento necessário para a predição determinística do escoamento através do meio poroso. Nesta tese é empregado um modelo lognormal para o campo de permeabilidades a fim de reproduzir-se a distribuição de permeabilidades do meio real, e a geração numérica destes campos aleatórios é feita pelo método da Soma Sucessiva de Campos Gaussianos Independentes (SSCGI). O objetivo principal deste trabalho é o estudo da quantificação de incertezas para o problema inverso do transporte de um traçador em um meio poroso heterogêneo empregando uma abordagem Bayesiana para a atualização dos campos de permeabilidades, baseada na medição dos valores da concentração espacial do traçador em tempos específicos. Um método do tipo Markov Chain Monte Carlo a dois estágios é utilizado na amostragem da distribuição de probabilidade a posteriori e a cadeia de Markov é construída a partir da reconstrução aleatória dos campos de permeabilidades. Na resolução do problema de pressão-velocidade que governa o escoamento empregase um método do tipo Elementos Finitos Mistos adequado para o cálculo acurado dos fluxos em campos de permeabilidades heterogêneos e uma abordagem Lagrangiana, o método Forward Integral Tracking (FIT), é utilizada na simulação numérica do problema do transporte do traçador. Resultados numéricos são obtidos e apresentados para um conjunto de realizações amostrais dos campos de permeabilidades.

Using sub-word-level information for confidence estimation with conditional random field models

The task of word-level confidence estimation (CE) for automatic speech recognition (ASR) systems stands to benefit from the combination of suitably defined input features from multiple information sources. However, the information sources of interest may not necessarily operate at the same level of granularity as the underlying ASR system. The research described here builds on previous work on confidence estimation for ASR systems using features extracted from word-level recognition lattices, by incorporating information at the sub-word level. Furthermore, the use of Conditional Random Fields (CRFs) with hidden states is investigated as a technique to combine information for word-level CE. Performance improvements are shown using the sub-word-level information in linear-chain CRFs with appropriately engineered feature functions, as well as when applying the hidden-state CRF model at the word level.

Bayesian Structured Prediction Using Gaussian Processes

We introduce a conceptually novel structured prediction model, GPstruct, which is kernelized, non-parametric and Bayesian, by design. We motivate the model with respect to existing approaches, among others, conditional random fields (CRFs), maximum margin Markov networks (M3N), and structured support vector machines (SVMstruct), which embody only a subset of its properties. We present an inference procedure based on Markov Chain Monte Carlo. The framework can be instantiated for a wide range of structured objects such as linear chains, trees, grids, and other general graphs. As a proof of concept, the model is benchmarked on several natural language processing tasks and a video gesture segmentation task involving a linear chain structure. We show prediction accuracies for GPstruct which are comparable to or exceeding those of CRFs and SVMstruct.

基于半马尔科夫条件随机场的命名体识别及其关系抽取研究

随着互联网和电子化办公的发展，出现了大量的文本资源。信息抽取技术可以帮助人们快速获取大规模文本中的有用信息。命名体识别与关系抽取是信息抽取的两个基本任务。本文在调研当前命名体识别和实体关系抽取中采用的主要方法的基础上，分别给出了解决方案。论文开展的主要工作有：（1）从模型选择和特征选择两个方面总结了命名体识别及实体关系抽取的国内外研究现状，重点介绍用于命名体识别的统计学习方法以及用于实体关系抽取的基于核的方法。（2）针对当前命名体识别中命名体片段边界的确定问题，研究了如何将 Semi-Markov CRFs 模型应用于中文命名体识别。这种模型只要求段间遵循马尔科夫规则，而段内的文本之间则可以被灵活的赋予各种规则。将这种模型用于中文命名体识别任务时，我们可以更有效更自由的设计出各种有利于识别出命名体片段边界的特征。实验表明，加入段相关的特征后，命名体识别的性能提高了 4-5 个百分点。（3）实体关系抽取的任务是判别两个实体之间的语义关系。之前的研究已经表明，待判别关系的两个实体间的语法树结构对于确定二者的关系类别是非常有用的，而相对成熟的基于平面特征的关系抽取方法在充分提取语法树结构特征方面的能力有限，因此，本文研究了基于核的中文实体关系抽取方法。针对中文特点，我们探讨了卷积（Convolution）核中使用不同的语法树对中文实体关系抽取性能的影响，构造了几种基于卷积核的复合核，改进了最短路依赖核。因为核方法开始被用于英文关系抽取时，F1 值也只有40%左右，而我们只使用作用在语法树上的卷积核时，中文关系抽取的F1 值达到了35%，可见核方法对中文关系抽取也是有效的。

Monte Carlo simulation of complex cohesive fracture in random heterogeneous quasi-brittle materials

A numerical method is developed to simulate complex two-dimensional crack propagation in quasi-brittle materials considering random heterogeneous fracture properties. Potential cracks are represented by pre-inserted cohesive elements with tension and shear softening constitutive laws modelled by spatially varying Weibull random fields. Monte Carlo simulations of a concrete specimen under uni-axial tension were carried out with extensive investigation of the effects of important numerical algorithms and material properties on numerical efficiency and stability, crack propagation processes and load-carrying capacities. It was found that the homogeneous model led to incorrect crack patterns and load–displacement curves with strong mesh-dependence, whereas the heterogeneous model predicted realistic, complicated fracture processes and load-carrying capacity of little mesh-dependence. Increasing the variance of the tensile strength random fields with increased heterogeneity led to reduction in the mean peak load and increase in the standard deviation. The developed method provides a simple but effective tool for assessment of structural reliability and calculation of characteristic material strength for structural design.

Learning discriminative sequence models from partially labelled data for activity recognition

Recognising daily activity patterns of people from low-level sensory data is an important problem. Traditional approaches typically rely on generative models such as the hidden Markov models and training on fully labelled data. While activity data can be readily acquired from pervasive sensors, e.g. in smart environments, providing manual labels to support fully supervised learning is often expensive. In this paper, we propose a new approach based on partially-supervised training of discriminative sequence models such as the conditional random field (CRF) and the maximum entropy Markov model (MEMM). We show that the approach can reduce labelling effort, and at the same time, provides us with the flexibility and accuracy of the discriminative framework. Our experimental results in the video surveillance domain illustrate that these models can perform better than their generative counterpart (i.e. the partially hidden Markov model), even when a substantial amount of labels are unavailable.