Gathering Information on the Web by Consistent Entity Augmentation

Table matching and data cleaning are two crucial activities in integrating data from different web tables, which have traditionally been considered as separate activities. We show that data cleaning can effectively help us discover table matches, and vice versa. In this paper, we study a hybrid machine-crowdsourcing approach to handle the two activities together with a well-developed knowledge base. Understanding the semantics of tables is fundamental to both matching and cleaning. We select the most valuable columns to crowdsourcing validation and infer others by consolidating crowdsourcing results and machine-generated results. When resolving inconsistency between data and semantics, relative trust is taken into account to validate data or semantics via crowd. Our experimental results show the effectiveness of the proposed approach for matching and cleaning web tables using real-life datasets.

Relation extraction from web tables contributes to a range of key semantic web applications including semantic table interpretation, RDF triplet extraction, and knowledge base augmentation. The structure of a table inherently provides high quality semantic relations between its columns. This paper introduces RelX, a novel, non exhaustive, fast converging algorithm for extracting column relations from web tables. RelX leverages existing linked data to infer relationships between table entities by observing their behaviours in the data. On carrying out preliminary evaluations, this technique yielded high quality and high accuracy against a manually annotated data-set.

Entity-oriented search has revolutionized search engines. In the era of Google Knowledge Graph and Microsoft Satori, users demand an effortless process of search. Whether they express an information need through a keyword query, expecting documents and entities, or through a clicked entity, expecting related entities, there is an inherent need for the combination of corpora and knowledge bases to obtain an answer. Such integration frequently relies on independent signals extracted from inverted indexes, and from quad indexes indirectly accessed through queries to a triplestore. However, relying on two separate representation models inhibits the effective cross-referencing of information, discarding otherwise available relations that could lead to a better ranking. Moreover, different retrieval tasks often demand separate implementations, although the problem is, at its core, the same. With the goal of harnessing all available information to optimize retrieval, we explore joint representation models of documents and entities, while taking a step towards the definition of a more general retrieval approach. Specifically, we propose that graphs should be used to incorporate explicit and implicit information derived from the relations between text found in corpora and entities found in knowledge bases. We also take advantage of this framework to elaborate a general model for entity-oriented search, proposing a universal ranking function for the tasks of ad hoc document retrieval (leveraging entities), ad hoc entity retrieval, and entity list completion. At a conceptual stage, we begin by proposing the graph-of-entity, based on the relations between combinations of term and entity nodes. We introduce the entity weight as the corresponding ranking function, relying on the idea of seed nodes for representing the query, either directly through term nodes, or based on the expansion to adjacent entity nodes. The score is computed based on a series of geodesic distances to the remaining nodes, providing a ranking for the documents (or entities) in the graph. In order to improve on the low scalability of the graph-of-entity, we then redesigned this model in a way that reduced the number of edges in relation to the number of nodes, by relying on the hypergraph data structure. The resulting model, which we called hypergraph-of-entity, is the main contribution of this thesis. The obtained reduction was achieved by replacing binary edges with n -ary relations based on sets of nodes and entities (undirected document hyperedges), sets of entities (undirected hyperedges, either based on cooccurrence or a grouping by semantic subject), and pairs of a set of terms and a set of one entity (directed hyperedges, mapping text to an object). We introduce the random walk score as the corresponding ranking function, relying on the same idea of seed nodes, similar to the entity weight in the graph-of-entity. Scoring based on this function is highly reliant on the structure of the hypergraph, which we call representation-driven retrieval. As such, we explore several extensions of the hypergraph-of-entity, including relations of synonymy, or contextual similarity, as well as different weighting functions per node and hyperedge type. We also propose TF-bins as a discretization for representing term frequency in the hypergraph-of-entity. For the random walk score, we propose and explore several parameters, including length and repeats, with or without seed node expansion, direction, or weights, and with or without a certain degree of node and/or hyperedge fatigue, a concept that we also propose. For evaluation, we took advantage of TREC 2017 OpenSearch track, which relied on an online evaluation process based on the Living Labs API, and we also participated in TREC 2018 Common Core track, which was based on the newly introduced TREC Washington Post Corpus. Our main experiments were supported on the INEX 2009 Wikipedia collection, which proved to be a fundamental test collection for assessing retrieval effectiveness across multiple tasks. At first, our experiments solely focused on ad hoc document retrieval, ensuring that the model performed adequately for a classical task. We then expanded the work to cover all three entity-oriented search tasks. Results supported the viability of a general retrieval model, opening novel challenges in information retrieval, and proposing a new path towards generality in this area.

Users often need to gather information about entities of interest. Recent efforts try to automate this task by leveraging the vast corpus of HTML tables; this is referred to as augmentation. The accuracy of entity augmentation critically depends on semantic relationships between web tables as well as semantic labels of those tables. Current techniques work well for string-valued and static attributes but perform poorly for numeric and time-varying attributes.In this paper, we first build a semantic graph that (i) labels columns with unit, scale and timestamp information and (ii) computes semantic matches between columns even when the same numeric attribute is expressed in different units or scales. Second, we develop a novel entity augmentation API suited for numeric and time-varying attributes that leverages the semantic graph. Building the graph is challenging as such label information is often missing from the column headers. Our key insight is to leverage the wealth of tables on the web and infer label information from semantically matching columns of other web tables; this complements local extraction from column headers. However, this creates an interdependence between labels and semantic matches; we address this challenge by representing the task as a probabilistic graphical model that jointly discovers labels and semantic matches over all columns. Our experiments on real-life datasets show that (i) our semantic graph contains higher quality labels and semantic matches and (ii) entity augmentation based on the above graph has significantly higher precision and recall compared with the state-of-the-art.

Topic models, which factor each document into different topics and represent each topic as a distribution of terms, have been widely and successfully used to better understand collections of text documents. However, documents are also associated with further information, such as the set of real-world entities mentioned in them. For example, news articles are usually related to several people, organizations, countries or locations. Since those associated entities carry rich information, it is highly desirable to build more expressive, entity-based topic models, which can capture the term distributions for each topic, each entity, as well as each topic-entity pair. In this paper, we therefore introduce a novel Entity Topic Model (ETM) for documents that are associated with a set of entities. ETM not only models the generative process of a term given its topic and entity information, but also models the correlation of entity term distributions and topic term distributions. A Gibbs sampling-based algorithm is proposed to learn the model. Experiments on real datasets demonstrate the effectiveness of our approach over several state-of-the-art baselines.

Information extraction from semi-structured webpages provides valuable long-tailed facts for augmenting knowledge graph. Relational Web tables are a critical component containing additional entities and attributes of rich and diverse knowledge. However, extracting knowledge from relational tables is challenging because of sparse contextual information. Existing work linearize table cells and heavily rely on modifying deep language models such as BERT which only captures related cells information in the same table. In this work, we propose a novel relational table representation learning approach considering both the intra- and inter-table contextual information. On one hand, the proposed Table Convolutional Network model employs the attention mechanism to adaptively focus on the most informative intra-table cells of the same row or column; and, on the other hand, it aggregates inter-table contextual information from various types of implicit connections between cells across different tables. Specifically, we propose three novel aggregation modules for (i) cells of the same value, (ii) cells of the same schema position, and (iii) cells linked to the same page topic. We further devise a supervised multi-task training objective for jointly predicting column type and pairwise column relation, as well as a table cell recovery objective for pre-training. Experiments on real Web table datasets demonstrate our method can outperform competitive baselines by +4.8% of F1 for column type prediction and by +4.1% of F1 for pairwise column relation prediction.

The qualitative and quantitative comparison of pre- and postoperative image data is an important possibility to validate surgical procedures, in particular, if computer assisted planning and/or navigation is performed. Due to deformations after surgery, partially caused by the removal of tissue, a non-rigid registration scheme is a prerequisite for a precise comparison. Interactive landmark-based schemes are a suitable approach, if high accuracy and reliability is difficult to achieve by automatic registration approaches. Incorporation of a priori knowledge about the anatomical structures to be registered may help to reduce interaction time and improve accuracy. Concerning pre- and postoperative CT data of oncological liver resections the intrahepatic vessels are suitable anatomical structures. In addition to using branching landmarks for registration, we here introduce quasi landmarks at vessel segments with high localization precision perpendicular to the vessels and low precision along the vessels. A comparison of interpolating thin-plate splines (TPS), interpolating Gaussian elastic body splines (GEBS) and approximating GEBS on landmarks at vessel branchings as well as approximating GEBS on the introduced vessel segment landmarks is performed. It turns out that the segment landmarks provide registration accuracies as good as branching landmarks and can improve accuracy if combined with branching landmarks. For a low number of landmarks segment landmarks are even superior.

Training of neural network can be accelerated by limited numerical precision together with specialized low-precision hardware. This paper studies how low precision can impact on entire training of RNNs. We emulate low precision training for recently proposed gated recurrent unit (GRU) and use dynamic fixed point as a target numeric format. We first show that batch normalization on input sequences can help speed up training with low precision as well as high precision. We also show that the overflow rate should be carefully controlled for dynamic fixed point. We study low precision training with various rounding options including bit truncation, round to nearest, and stochastic rounding. Stochastic rounding shows superior results than the other options. The effect of fully low precision training is also analyzed by comparing partial low precision training. We show that the piecewise linear activation function with stochastic rounding can achieve comparable training results with floating point precision. Low precision multiplier and accumulator (MAC) with linear-feedback shift register (LFSR) is implemented with 28nm Synopsys PDK for energy and performance analysis. Implementation results show low precision hardware is 4.7x faster, and energy per task is up to 4.55x lower than that of floating point hardware.

The energy problem in today’s society is becoming increasingly prominent, and the smart grid has become one of the important ways to solve the energy problem. Smart grid energy storage capacity planning and scheduling optimization is an important issue in the smart grid, which can make the grid more efficient, reliable, and sustainable to meet energy demand better and protect the environment. The core of smart grid energy storage capacity planning and scheduling optimization is maximizing the use of energy storage devices to balance the difference between power supply and demand to ensure the grid operation’s stability. Traditional planning methods are usually based on experience and rules, have low precision, and cannot adapt to the dynamic changes in the long-term development of the power grid. Therefore, this paper proposes a method that combines PSO-GRU (particle swarm Optimization (PSO)-gated recurrent unit (GRU)) and Multihead-Attention to realize smart grid energy storage capacity planning. And scheduling optimization. First, PSO-GRU models and predicts power grid data by searching for the optimal GRU model parameters; second, Multihead-Attention improves the model’s performance through the self-attention mechanism. Finally, we use the method to determine the optimal energy storage capacity and dispatching scheme for the efficient operation of smart grids. Our experiments use real power grid datasets and compare them with other common methods. Experimental results show that our proposed method has higher accuracy and stability than other methods and can better adapt to the dynamic changes of the power grid. This indicates that our method has good feasibility and applicability in practical applications and is significant for realizing the efficient operation of smart grids and energy saving and emission reduction.

Traditional two-dimensional vision only has basic pixel information and no depth information, while three-dimensional vision has depth information, which has high application value in the era of intelligence. The mainstream methods for obtaining depth information of objects include binocular stereo matching and TOF (Time of flight) cameras. The depth obtained by binocular stereo matching is with high precision in high-texture objects but with low precision in low-texture and repeated-texture objects, while the precision of depth obtained by TOF is the exact opposite. By combining binocular stereo matching and TOF, high-precision depth maps can be obtained in various scenarios. This paper proposes a new method of combining binocular stereo matching and TOF technology. The credible depth information is obtained through TOF to guide the stereo matching of the binocular camera, and then the information obtained after the guided matching is fused with the TOF information. The final fusion result is a high-precision depth map. The experimental results show that the fused depth map has higher accuracy and the execution speed within a reasonable range.

Under the laser tracker measurement of angle subsystem environment,signal acquisition system of laser tracker incremental encoder based on FPGA was designed in this paper. The FPGA in this system was mainly responsible for the encoder signal acquisition, processing, and generating timing signals, data caching mechanism and asynchronous output of the data. This paper detailedly introduced quadrupler frequency, phase discrimination, counting, latches, and UART asynchronous serial port output sub-modules in the process of signal acquisition of FPGA, which was implemented by Verilog HDL language. Online logic analyzer of the Quartus software acquired the value of encoder signal in real time, and the rationality of the design was verified ultimately.

Machine-readable semantic knowledge in the form of taxonomies (i.e., a collection of is-a edges) has proved to be beneficial in an array of NLP tasks including inference, textual entailment, question answering and information extraction. Such widespread utility of taxonomies has led to multiple large-scale manual efforts towards taxonomy induction such as WordNet and Cyc. However, manual construction of taxonomies is time-intensive, and usually, requires substantial annotation efforts by domain experts. Furthermore, the resulting taxonomies suffer from low coverage and are unavailable for specific domains or languages. Therefore, in recent years, there has been a growing body of work, which aims to induce taxonomies automatically, either from semi-structured knowledge resources or unstructured text. In this thesis, we focus on the task of automated taxonomy induction under a variety of different settings. We first focus on taxonomy induction from the largest semi-structured knowledge resource, i.e., Wikipedia. More specifically, we introduce a set of novel heuristics aimed towards inducing a large-scale taxonomy from the English Wikipedia categories network. We also propose a novel comprehensive path-based evaluation framework for taxonomies. Taxonomy induced using our approach significantly outperforms the state of the art across both edge-based as well as path-based evaluation metrics. Moreover, our experiments also demonstrate that high accuracy in edge-based evaluation metrics does not always translate to high-accuracy in path-based evaluation metrics. Subsequently, we propose a novel approach, which leverages the interlanguage links of Wikipedia to induce taxonomies in other Wikipedia languages. Compared to the state of the art, our approach is simpler, more principled, and results in taxonomies that are significantly more accurate across both edge-based and path-based evaluation metrics. In the second part of this thesis, we focus on taxonomy induction from unstructured text. We propose a novel approach towards taxonomy induction from an input vocabulary of seed terms. Unlike previous approaches, which extract singular hypernym edges for terms, our approach utilizes a novel probabilistic framework to extract long-range hypernym subsequences. Taxonomy induction from the extracted subsequences is cast as an instance of the minimum-cost flow optimization problem on a carefully-designed flow network. Through experiments, we demonstrate that our approach outperforms the state of the art across four languages. We also show that our approach is robust to the presence of noise in the input vocabulary. Our approach facilitates the relaxation of many simplifying assumptions, which were employed by previous taxonomy induction approaches. As a result, our work serves to automate the process of taxonomy induction from unstructured text in the true sense. Finally, we introduce a task of discovering and generalizing lexicalized templates from the titles of Wikipedia entities. The experimental results on this task demonstrate that taxonomies, which perform better on our proposed path-based metrics, result in a more accurate set of generalizations for a given set of entities. In summary, this thesis proposes new approaches towards automated taxonomy induction. It improves upon the state of the art in a variety of different settings. It also relaxes many simplifying assumptions that limited the applicability of prior approaches.

Bluetooth Low Energy (BLE) based Wireless Indoor Localization System (WILS) with high localization accuracy and high localization precision is a key requirement in enabling the Internet of Things (IoT) in today’s applications. In this paper, we investigated the effect of BLE signal variations on indoor localization caused by the change in BLE transmission power levels. This issue is not often discussed as most of the works on localization algorithms use the highest power levels but has important practical implications for energy efficiency, e.g., if a designer would like to trade-off localization performance and node lifetime. To analyze the impact, we used the established trilateration based localization model with two methods i.e., Centroid Approximation (CA) and Minimum Mean Square Error (MMSE). We observed that trilateration based localization with MMSE method outperforms the CA method. We further investigated the use of two filters i.e., Low Pass Filter (LPF) and Kalman Filter (KF) and evaluated their effects in terms of mitigating the random variations from BLE signal. In comparison to non-filter based approach, we observed a great improvement in localization accuracy and localization precision with a filter-based approach. Furthermore, in comparison to LPF based trilateration localization with CA, the performance of a KF based trilateration localization with MMSE is far better. An average of 1 m improvement in localization accuracy and approximately 50% improvement in localization precision is observed by using KF in trilateration based localization model with the MMSE method. In conclusion, with KF in trilateration based localization model with MMSE method effectively eliminates random variations in BLE RSS with multiple transmission power levels and thus results in a BLE based WILS with high accuracy and high precision.

In the digital signal processing, it is necessary to perform conversion from high quantization precision to low quantization precision of the intermediate signals, i.e. the truncation of the data bit width, due to the calculation precision of some digital algorithms. The direct truncation of the high quantization precision signals will lead to a decline in the spurious-free dynamic range (SFDR) of the truncated low quantization precision signals. Truncation after adding the Dither to the high quantization precision signals could effectively reduce the harmonic distortion of direct truncation error and significantly improve the dynamic range of the low quantization precision signals. The simulation results of MATLAB show that truncation after adding the Dither could increase the SFDR of the low quantization precision signals by about 16dB.

High control precision of pressure is one of the most important factor to guarantee the quality and quantity of synthetic diamond. But the pressure control system of cubic press in common use has the disadvantages of large zero drift, poor repeatability, and low control precision. To handle these problems, a novel pressure control system of cubic press has been developed. In this system, a cost-effective microprocessor ARM7TDMI (S3C44B0X) was employed, and an appropriate temperature compensation circuit was designed to reduce the negative influence of large environment temperature variation. The pressure boosting speed and pressure control precision were improved by three motors coordinated control. An efficient method was adopted to overcome the problems of large zero drift and poor repeatability by software. Furthermore, several measures were utilized to prevent the malfunction of system. Application results indicate that this control system has the advantages of low zero drift, high precision, good repeatability and reliability.