Fine-grained Factual Consistency Assessment for Abstractive Summarization Models

This PhD thesis combines two of the most exciting research areas of the last decades: quantum computing and machine learning. We introduce dissipative quantum neural networks (DQNNs), which are designed for fully quantum learning tasks, are capable of universal quantum computation and have low memory requirements while training. These networks are optimised with training data pairs in form of input and desired output states and therefore can be used for characterising unknown or untrusted quantum devices. We not only demonstrate the generalisation behaviour of DQNNs using classical simulations, but also implement them successfully on actual quantum computers. To understand the ultimate limits for such quantum machine learning methods, we discuss the quantum no free lunch theorem, which describes a bound on the probability that a quantum device, which can be modelled as a unitary process and is optimised with quantum examples, gives an incorrect output for a random input. Moreover we expand the area of applications of DQNNs in two directions. In the first case, we include additional information beyond just the training data pairs: since quantum devices are always structured, the resulting data is always structured as well. We modify the DQNN's training algorithm such that knowledge about the graph-structure of the training data pairs is included in the training process and show that this can lead to better generalisation behaviour. Both the original DQNN and the DQNN including graph structure are trained with data pairs in order to characterise an underlying relation. However, in the second extension of the algorithm we aim to learn characteristics of a set of quantum states in order to extend it to quantum states which have similar properties. Therefore we build a generative adversarial model where two DQNNs, called the generator and discriminator, are trained in a competitive way.

This paper presents a method for automatic summarization of JBIG2 coded textual images without optical character recognition (OCR). Compressed images are partially (less than 10% of the uncompressed image size) decompressed and text lines and words are marked. A few features are computed at each sentence level. Based on the feature values sentences are then marked as a summary sentence or not. The system finally generates a set of sentences as summary. In addition, sentences are ranked within the summary. Experiment considers Indian language text images. Test results show a sentence selection efficiency of about 56% when judged against summarization generated by human. A nonparametric (distribution-free) rank statistic shows a correlation coefficient of 0.28 as a measure of the (minimum) strength of the associations between sentence ranking by machine and human.

This article describes a density ratio approach to integrating external Language Models (LMs) into end-to-end models for Automatic Speech Recognition (ASR). Applied to a Recurrent Neural Network Transducer (RNN-T) ASR model trained on a given domain, a matched in-domain RNN-LM, and a target domain RNN-LM, the proposed method uses Bayes' Rule to define RNN-T posteriors for the target domain, in a manner directly analogous to the classic hybrid model for ASR based on Deep Neural Networks (DNNs) or LSTMs in the Hidden Markov Model (HMM) framework (Bourlard & Morgan, 1994). The proposed approach is evaluated in cross-domain and limited-data scenarios, for which a significant amount of target domain text data is used for LM training, but only limited (or no) {audio, transcript} training data pairs are used to train the RNN-T. Specifically, an RNN-T model trained on paired audio & transcript data from YouTube is evaluated for its ability to generalize to Voice Search data. The Density Ratio method was found to consistently outperform the dominant approach to LM and end-to-end ASR integration, Shallow Fusion.

Most current image captioning systems focus on describing general image content, and lack background knowledge to deeply understand the image, such as exact named entities or concrete events. In this work, we focus on the entity-aware news image captioning task which aims to generate informative captions by leveraging the associated news articles to provide background knowledge about the target image. However, due to the length of news articles, previous works only employ news articles at the coarse article or sentence level, which are not fine-grained enough to refine relevant events and choose named entities accurately. To overcome these limitations, we propose an Information Concentrated Entity-aware news image CAPtioning (ICECAP) model, which progressively concentrates on relevant textual information within the corresponding news article from the sentence level to the word level. Our model first creates coarse concentration on relevant sentences using a cross-modality retrieval model and then generates captions by further concentrating on relevant words within the sentences. Extensive experiments on both BreakingNews and GoodNews datasets demonstrate the effectiveness of our proposed method, which outperforms other state-of-the-arts. The code of ICECAP is publicly available at https://github.com/HAWLYQ/ICECAP.

Previous research in novelty detection has focused on the task of finding novel material, given a set or stream of documents on a certain topic. This study investigates the more difficult two-part task defined by the TREC 2002 novelty track: given a topic and a group of documents relevant to that topic, 1) find the relevant sentences from the documents, and 2) find the novel sentences from the collection of relevant sentences. Our research shows that the former step appears to be the more difficult part of this task, and that the performance of novelty measures is very sensitive to the presence of non-relevant sentences.

Photovoltaic (PV) power generation is highly intermittent in nature and any accurate very short-term prediction can decrease the impact of its uncertainties and operation costs and boost the reliable and efficient integration of PV systems into micro/smart grids. This work develops a new generalized technique for very short-term prediction of PV power generation from the lagged power generation data using fuzzy techniques. A preprocessor extracts relevant statistical features from the PV data which are fed to the fuzzy predictor. A modified version of Wang-Mendel training algorithm is employed to directly extract the fuzzy rules from the training data pairs. This methodology exploits the limited training data more efficiently. In addition, an online additive learning routine is proposed, which enables the predictor to learn from new data while running the predictions. So, the prediction accuracy increases over time and the predictor updates to account for long-term changing conditions of weather and PV system performance and its surroundings. Numerical results of the comparison of the proposed approach with simple fuzzy and traditional artificial neural network methods on a live PV system in the United Kingdom demonstrate its improved prediction accuracy, outperforming the benchmark approaches with a normalized mean absolute error (NMAE) of 3.6%.

Documents discussing public affairs, common themes, interesting products, and so on, are reported and distributed on the Web. Positive and negative opinions embedded in documents are useful references and feedbacks for governments to improve their services, for companies to market their products, and for customers to purchase their objects. Web opinion mining aims to extract, summarize, and track various aspects of subjective information on the Web. Mining subjective information enables traditional information retrieval (IR) systems to retrieve more data from human viewpoints and provide information with finer granularity. Opinion extraction identifies opinion holders, extracts the relevant opinion sentences, and decides their polarities. Opinion summarization recognizes the major events embedded in documents and summarizes the supportive and the nonsupportive evidence. Opinion tracking captures subjective information from various genres and monitors the developments of opinions from spatial and temporal dimensions. To demonstrate and evaluate the proposed opinion mining algorithms, news and bloggers' articles are adopted. Documents in the evaluation corpora are tagged in different granularities from words, sentences to documents. In the experiments, positive and negative sentiment words and their weights are mined on the basis of Chinese word structures. The f‐measure is 73.18% and 63.75% for verbs and nouns, respectively. Utilizing the sentiment words mined together with topical words, we achieve f‐measure 62.16% at the sentence level and 74.37% at the document level.

This work extends and evaluates a two-dimensional automatic evaluation metric for machine translation, which is designed to operate at the sentence level. The metric is based on the concepts of adequacy and fluency, aiming at decoupling both semantic and syntactic components of the translation process to provide a more balanced view on translation quality. These two elements are independently evaluated by using continuous space and <formula formulatype="inline" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex Notation="TeX">$n$</tex> </formula> -gram language modeling frameworks, respectively. Two different implementations are evaluated: a monolingual version that fully operates on the target language side, and a cross-language version that has the main advantage of not requiring reference translations. Both implementations are evaluated by comparing their performance with state-of-the-art automatic metrics over a dataset involving five different European languages.

&amp;lt;p&amp;gt;The current and future anthropogenic-induced high-latitude warming will have global climatic implications due to polar ice mass loss, sea level rise and ocean circulation changes. However, uncertainty remains on future climate projections mainly due to an incomplete understanding of climate, cryosphere and carbon cycle feedback processes occurring at centennial to millennial- timescales. Progress can be achieved by exploring climate and environmental changes that occurred in the past. In the HOTCLIM project, we are studying past warm periods, also referred to as interglacials, which exhibit a polar warming comparable to that projected by 2100 due to specific combinations of orbital and CO&amp;lt;sub&amp;gt;2&amp;lt;/sub&amp;gt; forcing. Especially, we are investigating the link between the carbon cycle dynamics and climate changes. To do so, we are combining (i) new analyses on the air trapped in Antarctic deep ice cores to inform on past changes in Antarctic climate and atmospheric CO2 concentrations (ii) climate and environmental data synthesis looking into the lower latitudes using terrestrial and oceanic archives (sea surface temperature, hydrological cycles, ocean circulation) (iii) an evaluation of outputs from climate models using the new comparison of the paleoclimatic datasynthesis and models output. The HOTCLIM project will improve our understanding of the natural climate variability and the processes involved during past periods associated with temperature changes comparable to projected future warming, hence helping improve climate projections&amp;lt;/p&amp;gt;&amp;lt;p&amp;gt;Here, we present the first results from the HOTCLIM project which is a multi-archive synthesis focused on the warm interval occurring between 190 and 243 ka BP, also refered to as Marine Isotopic Stage 7 (MIS 7). This warm period is of special interest because it follows the fastest transition between a cold (glacial) and a hot (interglacial) period of the last 800&amp;amp;#160;000 ka, with a polar warming of 10 degrees in less than 5ka. We have compiled more than 30 oceanic cores, 9 speleothems and 3 ice cores covering the MIS 7 period. To compare them, we are now building a common chronology to these records. The use of combined continental (ice cores, speleothems) and oceanic (sediment cores) archives located on the whole surface of the Earth will allows to characterize (i) the amplitude and the temporal structure of the surface warming across the globe (ii) the contrast between oceanic and continental warming.&amp;lt;/p&amp;gt;

Epidemiological surveillance for many important wildlife diseases relies on samples obtained from hunter-harvested animals. Statistical methods used to calculate sample size requirements assume that the target population is randomly sampled, and therefore the samples are representative of the population. But hunter-harvested samples may not be representative of the population due to disease distribution heterogeneities (e.g. spatial clustering of infected individuals), and harvest-related non-random processes like regulations, hunter selectivity, variable land access, and uneven hunter distribution. Consequently, sample sizes necessary for detection of disease are underestimated and disease detection probabilities are overestimated, resulting in erroneous inferences about disease presence and distribution.We have developed a modeling framework to support the design of efficient disease surveillance programs for wildlife populations. The constituent agent-based models can incorporate real-world heterogeneities associated with disease distribution, harvest, and harvest-based sampling, and can be used to determine population-specific sample sizes necessary for prompt detection of important wildlife diseases like chronic wasting disease and bovine tuberculosis. The modeling framework and its application has been described in detail by Belsare et al. [1]. Here we describe how model scenarios were developed and implemented, and how model outputs were analyzed. The main objectives of this methods paper are to provide users the opportunity to a) assess the reproducibility of the published model results, b) gain an in-depth understanding of model analysis, and c) facilitate adaptation of this modeling framework to other regions and other wildlife disease systems.•The two agent-based models, MOOvPOP and MOOvPOPsurveillance, incorporate real-world heterogeneities underpinned by host characteristics, disease spread dynamics, and sampling biases in hunter-harvested deer.•The modeling framework facilitates iterative analysis of locally relevant disease surveillance scenarios, thereby facilitating sample size calculations for prompt and reliable detection of important wildlife diseases.•Insights gained from modeling studies can be used to inform the design of effective wildlife disease surveillance strategies.

PermaBN: A Bayesian Network framework to help predict permafrost thaw in the Arctic

Complexity of data with respect to a particular class of neural networks is studied. Data complexity is measured by the magnitude of a certain norm of either the regression function induced by a probability measure describing the data or a function interpolating a sample of input/output pairs of training data chosen with respect to this probability. The norm is tailored to a type of computational units in the network class. It is shown that for data for which this norm is “small”, convergence of infima of error functionals over networks with increasing number of hidden units to the global minima is relatively fast. Thus for such data, networks with a reasonable model complexity can achieve good performance during learning. For perceptron networks, the relationship between data complexity, data dimensionality and smoothness is investigated.KeywordsProbability MeasureRegression FunctionData ComplexityHide UnitComputational UnitThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Degraded Structure and Hue Guided Auxiliary Learning for low-light image enhancement

The Massachusetts Institute of Technology general circulation model is used for the analysis of baroclinic tides over Anton Dohrn Seamount (ADS), in the North Atlantic. The model output is validated against in situ data collected during the 136th cruise of the RRS “James Cook” in May–June 2016. The observational data set includes velocity time series recorded at two moorings as well as temperature, salinity, and velocity profiles collected at 22 hydrological stations. Synthesis of observational and model data enabled the reconstruction of the details of baroclinic tidal dynamics over ADS. It was found that the baroclinic tidal waves are generated in the form of tidal beams radiating from the ADS periphery to its center, focusing tidal energy in a surface layer over the seamount's summit. This energy focusing enhances subsurface water mixing and the local generation of internal waves. The tidal beams interacting with the seasonal pycnocline generate short‐scale internal waves radiating from the ADS center. An important ecological outcome from this study concerns the pattern of residual currents generated by tides. The rectified flows over ADS have the form of a pair of dipoles, cyclonic and anticyclonic eddies located at the seamount's periphery. These eddies are potentially an important factor in local larvae dispersion and their escape from ADS.

Supervised super-resolution methods estimate high-resolution simulation results from low-resolution inputs. These methods typically train the neural networks using pairs of high-resolution data and their down-sampled low-resolution counterparts, implicitly assuming the global similarity between the low-resolution and high-resolution models. This assumption, however, often fails as the low-resolution models usually exhibit stiffer behavior than high-resolution models due to numerical stiffening. This paper proposes a novel supervised super-resolution method to mitigate this problem for linear deformable object simulation. The method constructs training data pairs by matching low-resolution and high-resolution simulation snapshots based on the similarity of the normalized strain energy, the normalized temporal change rate of strain energy, and the modal contributions to the overall strain energy. The loss function also incorporates the equation residuals derived from the finite element method. The time integration scheme is selected by examining the eigenvalue distributions of the neural tangent kernel associated with the finite element residuals. Compared with previous down-sampling methods, the proposed method reduces the maximum relative displacement error by 25%, 54% and 38% for the beam, elephant and armadillo models, respectively.