Advances in Artificial Intelligence: From Theory to Practice

This paper aims at the annotation of movement phrases in Vietnamese folk dance videos that were mainly gathered, stored and used in teaching at art schools and in preserving cultural intangible heritages (performed by different famous folk dance masters). We propose a framework of automatic movement phrase annotation, in which the motion vectors are used as movement phrase features. Movement phrase classification can be carried out, based on dancer’s trajectories. A deep investigation of Vietnamese folk dance gives an idea of using optical flow as movement phrase features in movement phrase detection and classification. For the richness and usefulness in annotation of Vietnamese folk dance, a lookup table of movement phrase descriptions is defined. In initial experiments, a sample movement phrase dataset is built up to train k-NN classification model. Experiments have shown the effectiveness of the proposed framework of automatic movement phrase annotation with classification accuracy at least 88%.

Analysis of behavioral data represents today a big issue, as so many domains generate huge quantity of activity and mobility traces. When traces are labeled by the user that generates it, models can be learned to accurately predict the user of an unknown trace. In online systems however, users may have several virtual identities, or duplicate labels. By ignoring them, the prediction accuracy drastically drops, as the set of all virtual identities of a single person is not known beforehand. In this article, we tackle this duplicate labels identification problem, and present an original approach that explores the lattice of binary classifiers. Each subset of labels is learned as the positive class against the others (the negative class), and constraints make possible to identify duplicate labels while pruning the search space. We experiment this original approach with data of the video game Starcraft 2 in the new context of Electronic Sports (eSport) with encouraging results.

In Southeast Asia, Thai dance is a living traditional art form that belongs to the Intangible Cultural Heritage listed by UNESCO. This unique and stylized traditional dance portrays its history, culture, emotional expression, body movement etc. To archive the knowledge of the Traditional Thai dance, a dance notation known as “Labanotation” has been widely used to record and archive the unique essence of dance knowledge. In addition, many researchers have worked on reproducing and showcasing the dance movements. Currently, there is no such system that is available for us to demonstrate our Thai Dance Notation Score in 3D animation. Specifically, displaying the hand and finger movement is an issue. The aim of this paper is to present the process of implementing a tool to translate the dance notation into a 3D animation focusing on hand and finger movements of the traditional Thai dance.

In this preliminary work, we propose a system prototype for Thai Dance training. This paper considers the problem of teaching traditional dances from Thailand. This is particularly useful given the lack of teachers and tools for teaching dances. In order to build a software tool helping people learn Thai dances, the main problems are (i) how to represent the dance gestures and movements of the dance to teach, (ii) how to display it for the learner and how to rate the performance of the learner and provide him useful feedback. Fortunately, Natural User Interfaces (NUI) enables users to interact with a system in a natural and intuitive way. For instance, a user can interact with the system by his body through postures and movements. In this study, we developed a working prototype of a system teaching users traditional Thai dances. The system requires Kinect-based device to enable real-time skeleton tracking. For the reference postures/movements dataset, we collected dance movement from experts by Motion Capture System and used the collected data to represent the dance in the system. Moreover, the system is designed such that it rates the user’s performance and provides helpful and real-time feedback to the user.

Creating the behavior for non-player characters (NPCs) in video games is a complex task that requires the collaboration among programmers and game designers. Usually these game designers are responsible of configuring and fine tuning certain parameters of the behavior, while programmers write the actual code of those behaviors. That requires several iterations between them. In this paper, we present a new approach for creating the behavior of NPCs that gives more power to the game designer to create behavior without technical knowledge using program by demonstration but preserving the designer confident of the final behavior.

People cannot use a text search to find mathematical expressions because expressions cannot be replaced with words. Our research uses an ordinary text search and presents appropriate mathematical expression images (hereinafter called math-images) for input keywords. First we classify a set of the top ranking images from all the images in HTML files by scoring them. We focus on three viewpoints that are unique to mathematical expression images and mark the images by using these viewpoints. Then by adding bonus points to these marked images, the best three images are chosen from the set and presented with an explanation of the keyword and the surrounding information in the HTML files. We conducted two experiments to optimize the parameters of the expression giving the mark and to evaluate the effect of the bonus points. The rate of the average correct images of the best three was 79.5%.

Chiang Mai is known as the capital city of handicraft and tourism industry of Thailand. It generates source of income and employment to the local people for a long time. Nowadays, the craftsmen have tried to exploited their wisdom, creativity, skills and technology to enhance the product’s value in order to create the sustainable and higher economic growth. However, most of the time, the created value was not delivered along with the product or transferred via seller. Therefore, buyer could not perceive the wisdom, creativity, or craftsmanship skill of the purchased product. Our work aimed at applying the notion of digital content e.g. storytelling, 360-degree images, 3D-model, or augmented reality to preserve the added value and deliver to the customer. Then, the digital content was conveyed to customer through various IT tools i.e. website, web application, video streaming, and mobile application.

The aim of this work is to produce and test a robust, distributed, multi-agent task allocation algorithm, as these are scarce and not well-documented in the literature. The vehicle used to create the robust system is the Performance Impact algorithm (PI), as it has previously shown good performance. Three different variants of PI are designed to improve its robustness, each using Monte Carlo sampling to approximate Gaussian distributions. Variant A uses the expected value of the task completion times, variant B uses the worst-case scenario metric and variant C is a hybrid that implements a combination of these. The paper shows that, in simulated trials, baseline PI does not handle uncertainty well; the task-allocation success rate tends to decrease linearly as degree of uncertainty increases. Variant B demonstrates a worse performance and variant A improves the failure rate only slightly. However, in comparison, the hybrid variant C exhibits a very low failure rate, even under high uncertainty. Furthermore, it demonstrates a significantly better mean objective function value than the baseline.

We propose to use a recently introduced merging procedure for jointly inconsistent probabilistic assessments to the statistical matching problem. The merging procedure is based on an efficient L1 distance minimization through mixed-integer linear programming that results not only feasible but also meaningful for imprecise (lower-upper) probability evaluations elicitation. Significance of the method can be appreciated whenever among quantities (events) there are logical (structural) constraints and there are different sources of information. Statistical matching problem has these features and is characterized by a set of random (discrete) variables that cannot be jointly observed. Separate observations share some common variable and this, together with structural constraints, make sometimes inconsistent the estimates of probability occurrences. Even though estimates on statistical matching are mainly conditional probabilities, inconsistencies appear only on events with the same conditioning, hence the correction procedure can be easily reduced to unconditional cases and the aforementioned procedure applied.

In probability theory the notion of coherence has been introduced by de Finetti in terms of bets and it reveals to be equivalent to the notion of consistence of a partial assessment with a finitely additive probability. An important feature of coherent assessments is their coherent extendibility: in general we obtain a class of coherent extensions, determining a lower and an upper envelope. A similar notion of coherence has been recently introduced for (T-conditional) possibility measures, where T is a t-norm. The extendibility of coherent possibility assessments reveals to be particularly suitable for studying interval-based possibilistic logic. Our aim is to compare the results implied by the coherent setting with those obtained in different approaches, in particular, that relying on classical T-based conditioning.

We generalize, in the setting of coherence, the notions of conjunction and disjunction of two conditional events to the case of n conditional events. Given a prevision assessment on the conjunction of two conditional events, we study the set of coherent extensions for the probabilities of the two conditional events. Then, we introduce by a progressive procedure the notions of conjunction and disjunction for n conditional events. Moreover, by defining the negation of conjunction and of disjunction, we show that De Morgan’s Laws still hold. We also show that the associative and commutative properties are satisfied. Finally, we examine in detail the conjunction for a family $$\mathcal F$$ of three conditional events. To study coherence of prevision for the conjunction of the three conditional events, we need to consider the coherence for the prevision assessment on each conditional event and on the conjunction of each pair of conditional events in $$\mathcal F$$.

Crowdsourcing, in particular microtasking is now a powerful concept used by employers in order to obtain answers on tasks hardly handled by automated computation. These answers are provided by human employees and then combined to get a final answer. Nevertheless, the quality of participants in microtasking platforms is often heterogeneous which makes results imperfect and thus not fully reliable. To tackle this problem, we propose a new approach of label aggregation based on gold standards under the belief function theory. This latter provides several tools able to represent and even combine imperfect information. Experiments conducted on both simulated and real world datasets show that our approach improves results quality even with a high ratio of bad workers.

De Finetti’s approach of an event of two levels of knowledge was recently proposed as the model of reference for psychology studies. We show that de Finetti’s qualitative probability framework seems to be “natural” to children aged from 3 to 4 as well as to account for children’s heuristic approach to probabilistic reasoning.

This short note studies a multiple source extension of categorical mass functions in the sense of Shafer’s evidence theory. Each subset of possible worlds is associated with a subset of information sources and represents a tentative description of what is known. Analogs of belief, plausibility, commonality functions, valued in terms of subsets of agents or sources, are defined, replacing summation by set union. Set-valued plausibility is nothing but set-valued possibility because it is union-decomposable with respect to the union of events. In a special case where each source refers to a single information item, set-valued belief functions decompose with respect to intersection and are thus multiple source necessity-like function. Connections with Belnap epistemic truth-values for handling multiple source inconsistent information are shown. A formal counterpart of Dempster rule of combination is defined and discussed as to its merits for information fusion.

WalkSAT is a local search algorithm conceived for solving SAT problems, which is also used for sampling possible worlds from a logical formula. This algorithm is used by Markov Logic Networks to perform slice sampling and give probabilities from a knowledge base defined with soft and hard constraints. In this paper, we will show that local search strategies, such as WalkSAT, may perform as poorly as a pure random walk on a category of problems that are quite common in industrial fields. We will also give some insights into the reasons that make random search algorithms intractable for these problems.

To support health monitoring and life-long capability management for self-sustaining manufacturing systems, next generation machine components are expected to embed sensory capabilities combined with advanced ICT. The combination of sensory capabilities and the use of Object-Oriented Bayesian Networks (OOBNs) supports self-diagnosis at the component level enabling them to become self-aware and support self-healing production systems. This paper describes the use of a modular component-based modelling approach enabled by the use of OOBNs for health monitoring and root-cause analysis of manufacturing systems using a welding controller produced by Harms & Wende (HWH) as an example. The model is integrated into the control software of the welding controller and deployed as a SelComp using the SelSus Architecture for diagnosis and predictive maintenance. The SelComp provides diagnosis and condition monitoring capabilities at the component level while the SelSus Architecture provides these capabilities at a wider system level. The results show significant potential of the solution developed.

A multiple-agent logic, which associates subsets of agents to logical formulas, has been recently proposed. The paper presents a graphical counterpart of this logic, based on a multiple agent version of possibilistic conditioning, and applies it to preference modeling. First, preferences of agents are supposed to be all or nothing. We discuss how one can move from the network to the logic representation and vice-versa. The new representation enables us to focus on networks associated to subsets of agents, and to identify inconsistent agents, or conflicting subsets of agents. The question of optimization and dominance queries is discussed. Finally, the paper outlines an extension where gradual preferences are handled.

Fault Trees or Bow Tie Diagrams are widely used for system dependability assessment. Some probabilistic extensions have been proposed by using Bayesian network formalism. This article proposes a general modeling approach under the form of a probabilistic relational model (PRM), relational extension of Bayesian networks, that can represent any fault tree, defined as an event tree with possible safety barriers, simply described in a relational database. We first describe an underlying relational schema describing a generic fault tree, and the probabilistic dependencies needed to model the existence of an event given the possible existence of its related causes and eventual safety barriers.

Evidential graphical models are considered as an efficient tool for representing and analyzing complex and real-world systems, and reasoning under uncertainty.This work raises the issue of estimating the different parameters of these networks. More precisely, we address the problem of updating these parameters when getting new data without repeating the learning process from the beginning. Indeed, we propose a new incremental approach to update the different parameters based on the combination rules proposed in the evidence framework.

This paper presents the aGrUM framework, a C++ library providing state-of-the-art implementations of graphical models for decision making, including Bayesian Networks, Influence Diagrams, Credal Networks, Probabilistic Relational Models. This is the result of an ongoing effort to build an efficient and well maintained open source cross-platform software, running on Linux, MacOS X and Windows, for dealing with graphical models. The framework also contains a wrapper, pyAgrum, for exploiting aGrUM within Python.

We propose a new approach to detect credit card fraud based on suspicious payment patterns. According to our hypothesis fraudsters use stolen credit card data at specific, recurring sets of shops. We exploit this behavior to identify fraudulent transactions. In a first step we show how suspicious patterns can be identified from known compromised cards. The transactions between cards and shops can be represented as a bipartite graph. We are interested in finding fully connected subgraphs containing mostly compromised cards, because such bicliques reveal suspicious payment patterns. Then we define new attributes which capture the suspiciousness of a transaction indicated by known suspicious patterns. Eventually a non-linear classifier is used to assess the predictive power gained through those new features. The new attributes lead to a significant performance improvement compared to state-of-the-art aggregated transaction features. Our results are verified on real transaction data provided by our industrial partner (Worldline http://www.worldline.com).

Air leakage in braking pipes is a commonly encountered mechanical defect on trains. A severe air leakage will lead to braking issues and therefore decrease the reliability and cause train delays or stranding. However, air leakage is difficult to be detected via visual inspection and therefore most air leakage defects are run to fail. In this study we present a contextual anomaly detection method that detects air leakage based on the on/off logs of a compressor. Air leakage causes failure in the context when the compressor idle time is short than the compressor run time, that is, the speed of air consumption is faster than air generation. In our method the logistic regression classifier is adopted to model two different classes of compressor behavior for each train separately. The logistic regression classifier defines the boundary separating the two classes under normal situations and models the distribution of the compressor idle time and run time separately using logistic functions. The air leakage anomaly is further detected in the context that when a compressor idle time is erroneously classified as a compressor run time. To distinguish anomalies from outliers and detect anomalies based on the severity degree, a density-based clustering method with a dynamic density threshold is developed for anomaly detection. The results have demonstrated that most air leakages can be detected one to four weeks before the braking failure and therefore can be prevented in time. Most importantly, the contextual anomaly detection method can pre-filter anomaly candidates and therefore avoid generating false alarms.

Detecting anomalies in the flow of system logs of a high performance computing (HPC) facility is a challenging task. Although previous research has been conducted to identify nominal and abnormal phases; practical ways to provide system administrators with a reduced set of the most useful messages to identify abnormal behaviour remains a challenge. In this paper we describe an extensive study of logs classification and anomaly detection using K-means on real HPC unlabelled data extracted from the Curie supercomputer. This method involves (1) classifying logs by format, which is a valuable information for admin, then (2) build normal and abnormal classes for anomaly detection. Our methodology shows good performances for clustering and detecting abnormal logs.

In Twitter based applications such as tweet summarization, the existence of ill-intentioned users so-called spammers imposes challenges to maintain high performance level in those applications. Conventional social spammer/spam detection methods require significant and unavoidable processing time, extending to months for treating large collections of tweets. Moreover, these methods are completely dependent on supervised learning approach to produce classification models, raising the need for ground truth data-set. In this paper, we design an unsupervised language model based method that performs collaboration with other social networks to detect spam tweets in large-scale topics (e.g. hashtags). We experiment our method on filtering more than 6 million tweets posted in 100 trending topics where Facebook social network is accounted in the collaboration. Experiments demonstrate highly competitive efficiency in regards to processing time and classification performance, compared to conventional spam tweet detection methods.

Predicting future river flow is a difficult problem. Firstly, models are (by definition) crudely simplified versions of reality. Secondly, historical streamflow data is limited and noisy. Bayesian model averaging is theoretically a good way to cope with these difficulties, but it has not been widely used on this and similar problems. This paper uses real-world data to illustrate why. Bayesian model averaging can give a better prediction, but only if the amount of data is small — if the data is consistent with a wide range of different models (instead of unambiguously consistent with only a narrow range of near-identical models), then the weighted votes of those diverse models will give a better prediction than the single best model. In contrast, with plenty of data, only a narrow range of near-identical models will fit that data, and they all vote the same way, so there is no improvement in the prediction. But even when the data supports a diverse range of models, the improvement is far from large, but it is the direction of the improvement that can predict more accurately. Working around these caveats lets us better predict floods and similar problems, using limited or noisy data.

Mixed fruit-vegetable cropping systems are a promising way of ensuring environmentally sustainable agricultural production systems in response to the challenge of being able to fulfill local market requirements. Indeed, they combine productions and they make a better use of biodiversity. These agroforestry systems are based on a complex set of interactions modifying the utilization of light, water and nutrients. Thus, designing such a system must optimize the use of these resources: by maximizing positive interactions (facilitation) and minimizing negative ones (competition). To attain these objectives, the system’s design has to include the spatial and temporal dimensions, taking into account the evolution of above- and belowground interactions over a time horizon. For that, we define the Mixed Fruit-Vegetable Crop Allocation Problem (MFVCAP) using a discrete representation of the land and the interactions between vegetable crops and fruit trees. First, we give a direct formulation as a binary quadratic program (BQP). Then we reformulate the problem using a Benders decomposition approach. The master problem has 0/1 binary variables and deals with tree positioning. The subproblem deals with crop quantities. The BQP objective function becomes linear in the continuous subproblem by exploiting the fact that it depends only on the quantity of crops assigned to land units having shade, root, or nothing. This problem decomposition allows us to reformulate the MFVCAP into a Mixed Integer linear Program (MIP). The detailed spatial-temporal crop allocation plan is easy to obtain after solving the MIP. Experimental results show the efficiency of our approach compared to a direct solving of the original BQP formulation.

The emergence of widely available connected devices is perceived as the promise of new added-value services. Companies can now gather, often in real time, huge amounts of data about their customers’ habits. Seemingly, all they have to do is to mine these raw data in order to discover the profiles of their users and their needs.Stemming from an industrial experience, this paper, however, shows that things are not that simple. It appears that, even in an exploratory data mining phase, the usual data cleaning and preprocessing steps are a long shot from being adequate. The rapid deployment of connected devices indeed introduces its own series of problems. The paper shares the pitfalls encountered in a project aiming at enhancing the cooking habits and presents some hard learnt lessons of general import.

This paper presents a method to assess nitrogen levels, a nitrogen nutrition index (NNI), in corn crops (Zea mays) using multispectral remote sensing imagery. The multispectral sensors used were four spectral bands only. The experiments were compared with nitrogen levels sensed in the field. The corn crops were divided into three nitrogen fertilization levels (70, 140 and 210 $$\mathrm{kg N}\cdot \mathrm{ha}^{-1}$$) into three replicates. In this sense, we propose a method to infer nitrogen levels in corn crops by using airborne multispectral sensors and machine learning techniques. The presented results offered a simple model to estimate nitrogen with low-cost technologies (UAVs and multispectral cameras only) in small to medium size areas of corn crops.

The paper introduces a method to evaluate a short food supply chain based on argumentation. It defines an analytical argumentation system using contexts, and introduces indicators to perform analysis. It proposes an evaluation of the experimental device created to observe the short food supply chain mechanisms, based on this analysis methodology. It concludes on the feedback learnt from this analysis, from methodological and application viewpoints.

Although modeling argument structures is helpful to make involved parties understand the pros and cons of an issue and the context of each other’s positions, stakeholders have no means to anticipate the impacts of adopting the debated solutions, let alone to compare them. This is where using simulation approaches would greatly enrich the deliberation process. This paper introduces an approach combining argumentation and simulation. We consider a case study in which both are used to assess and compare cultural options available to farmers.

In this paper we present an implemented method for analysing arguments from drug reviews given by patients in medical forums on the web. For this we provide a number of classification rules which allow for the extraction of specific arguments from the drug reviews. For each review we use the extracted arguments to instantiate a Dung argument graph. We undertake an evaluation of the resulting argument graphs by applying Dung’s grounded semantics. We demonstrate a correlation between the arguments in the grounded extension of the graph and the rating provided by the user for that particular drug.

A dynamic framework, based on the Dynamic Logic of Propositional Assignments ($$\mathsf {DL\text {-}PA}$$), has recently been proposed for Dung’s abstract argument system. This framework allows the addition and the removal of attacks, and the modification of the acceptance status of arguments. We here extend this framework in order to capture the addition and the removal of arguments. We then apply the framework on an access control case, where an agent engages in an argued dialogue to access some information controlled by another agent.

Analogical reasoning makes use of a kind of resemblance of one thing to another for assigning properties from one context to another. This kind of reasoning is used quite often by human beings, especially in unseen situations. The key idea of analogy is to identify a good similarity; however, similarity may be varied on subjective factors (i.e. an agent’s preferences). This paper studies an implementation of this phenomena using an answer set programming with Description Logics. The main idea underlying the proposed approach lies in the so-called Argument from Analogy developed by Walton [1]. Finally, the paper relates the approach to others and discusses future directions.

This paper uses argumentation as the basis for modeling and implementing the relevant legislation of an EU country relating to medical data access. Users can consult a web application for determining their allowed level of access to a patient’s medical record and are offered an explanation based on the relevant legislation. The system can also advise a user on what additional information is required for a higher access level. The system is currently in the process of an extensive evaluation through a pilot trial with a special focus group of medical professionals. The development methodology that we have used is generally applicable to any other similar cases of decision making based on legislative regulations. The main advantage of using argumentation is the ability to explain the solutions drawn and the high modularity of software facilitating the extension and adaptation of the system when new relevant legislation becomes available.

We present an approach to reasoning with knowledge bases comprised of strict and defeasible rules over literals. A controlled natural language is proposed as a human/machine interface to facilitate the specification of knowledge and verbalisation of results. Techniques from formal argumentation theory are employed to justify conclusions of the approach; this aims at facilitating human acceptance of computed answers.

Computational models of human processes are used for many different purposes and in many different types of applications. A common challenge in using such models is to find suitable parameter values. In many cases, the ideal parameter values are those that yield the most realistic simulation results. However, there are situations in which the goodness of fit is not the main or only criterion to evaluate the appropriateness of a model, but where other aspects of the model behavior are also relevant. This is often the case when computational models are employed in real-life applications, such as mHealth systems. In this paper, we explore how parameter tuning techniques can be used to analyze the behavior of computational models systematically and to investigate the reasons behind the observed behavior. We study a computational model of psychosocial influences on physical activity behavior as an in-depth use case. In this particular case, an important measure of the feasibility of the model is the diversity in the simulation outcomes. This novel application of parameter tuning techniques for analysis and understanding of model behavior is transferable to other cases, and is therefore a valuable new approach in the toolset of computational modelers.

The adoption and maintenance of a healthy lifestyle is a fundamental pillar in the quest towards a healthy society. Modern (mobile) technology allows for increasingly intelligent systems that can help to optimize people’s health outcomes. One of the possible directions in such mHealth systems is the use of intelligent reasoning engines based on dynamic computational models of behavior change. In this work, we investigate the accuracy of such a model to simulate changes in physical activity levels over a period of two to twelve weeks. The predictions of the model are compared to empirical physical activity data of 108 participants. The results reveal that the model’s predictions show a moderate to strong correlation with the actual data, and it performs substantially better than a simple alternative model. Even though the implications of these findings depend strongly on the application at hand, we show that it is possible to use a computational model to predict changes in behavior. This is an important finding for developers of mHealth systems, as it confirms the relevance of model-based reasoning in such health interventions.

One common health problem in the US faced by colleges and universities is binge drinking. College students often post drinking related texts and images on social media as a socially desirable identity. Some public health and clinical research scholars have surveyed different social media sites manually to understand their behavior patterns. In this paper, we investigate the feasibility of mining the heterogeneous data scattered on social media to identify drinking-related contents, which is the first step towards unleashing the potential of social media in automatic detection of binge drinking users. We use the state-of-the-art algorithms such as Support Vector Machine and neural networks to classify drinking from non-drinking posts, which contain not only text, but also images and videos. Our results show that combining heterogeneous data types, we are able to identify drinking related posts with an overall accuracy of 82%. Prediction models based on text data is more reliable compared to the other two models built on image and video data for predicting drinking related contents.

Data on disease burden is often used for assessing population health, evaluating the effectiveness of interventions, formulating health policies, and planning future resource allocation. We investigated whether Internet usage data, particularly the search volume on Google and page view counts on Wikipedia, are correlated with the disease burden, measured by prevalence and treatment cost, for 1,633 diseases over an 11-year period. We also applied the method of least absolute shrinkage and selection operator (LASSO) to predict the burden of diseases, using those Internet data together with three other variables we quantified previously. We found a relatively strong correlation for 39 of 1,633 diseases, including viral hepatitis, diabetes mellitus, other headache syndromes, multiple sclerosis, sleep apnea, hemorrhoids, and disaccharidase deficiency. However, an accurate analysis must consider each condition’s characteristics, including acute/chronic nature, severity, familiarity to the public, and presence of stigma.

Statistical and machine learning approaches to named entity recognition have risen to prominence in the field of natural language processing. Certain named entities, specifically biomedical software, is a challenge to identify as a named entity. One direction is investigating the use of contextual semantic information to assist in this task as alluded to by previous researchers. We introduce an ontology-driven method that experiments with both information extraction and inherited features of ontologies (e.g., embedded semantic relationships and links to entities) to automatically identify familiar and unfamiliar software names. We evaluated this method with a set of biomedical research abstracts containing software entities. Our proposed approach could be used to further augment other named entity recognition methods.

Mental health is an increasingly important problem in healthcare. Psychiatric stressors are one of the major contributors of mental disorders. Very few studies have investigated stressor data in electronic health records, mostly because they are recorded in narrative texts. This study takes the initiative to develop a natural language processing system to automatically extract psychiatric stressors from clinical notes. Our approach integrates domain knowledge from multiple sources and unsupervised word representation features generated from deep learning based algorithms, to address the context dependence and data sparseness challenges caused by idiosyncratic psychosocial backgrounds. Experimental results on psychiatric notes from the CEGS N-GRID 2016 challenge demonstrate that the proposed approach is promising. The best performing configuration achieved a precision of 90.5%, a recall of 65.5%, and a F-measure of 76.0% for inexact matching.

The benefits of crowdsourcing have been widely recognized in active learning for text mining. Due to the lack of golden ground-truth, it is crucial to evaluate how trustworthy of “noisy” labelers when labeling informative instances. Despite recent achievements made on active learning with crowdsourcing, most of the research works are involved in tuning a considerable amount of parameters, and also sensitive to noise. In this paper, a novel framework to select both the best-fitting labeler and the most informative instance is proposed, with the help of the minimum description length principle which is acknowledged as noise-tolerant and parameter-free. The algorithm is proved to be effective through extensive experiments on texts.

Lyrics take a great role to express users’ feelings. Every user has its own patterns and styles of songs. This paper proposes a method to capture the patterns and styles of users and generates lyrics automatically, using Long Short-Term Memory network combined with language model. The Long Short-Term memory network can capture long-term context information into the memory, this paper trains the context representation of each line of lyrics as a sentence vector. And with the recurrent neural network-based language model, lyrics can be generated automatically. Compared to the previous systems based on word frequency, melodies and templates which are hard to be built, the model in this paper is much easier and fully unsupervised. With this model, some patterns and styles can be seen in the generated lyrics of every single user.

Great efforts have been dedicated to harvesting knowledge bases from online encyclopedias. These knowledge bases play important roles in enabling machines to understand texts. However, most current knowledge bases are in English and non-English knowledge bases, especially Chinese ones, are still very rare. Many previous systems that extract knowledge from online encyclopedias, although are applicable for building a Chinese knowledge base, still suffer from two challenges. The first is that it requires great human efforts to construct an ontology and build a supervised knowledge extraction model. The second is that the update frequency of knowledge bases is very slow. To solve these challenges, we propose a never-ending Chinese Knowledge extraction system, CN-DBpedia, which can automatically generate a knowledge base that is of ever-increasing in size and constantly updated. Specially, we reduce the human costs by reusing the ontology of existing knowledge bases and building an end-to-end facts extraction model. We further propose a smart active update strategy to keep the freshness of our knowledge base with little human costs. The 164 million API calls of the published services justify the success of our system.

This paper aims at demonstrating sentiment strength analysis in aspect-based opinion mining. Previous works normally focused on reviewers’ sentiment orientation and ignored sentiment strength that users expressed in the reviews. In order to offset this disadvantage, two methods for sentiment strength evaluation were proposed. Experiments on a huge hotel review dataset show how sentiment strength analysis can improve the performance of aspect rating prediction.

Multi-document summarization is a difficult natural language processing task. Many extractive summarization methods consist of two steps: extract important concepts of documents and select sentences based on those concepts. In this paper, we introduce a method to use the Latent Dirichlet Allocation (LDA) topic labels as concepts, instead of n-gram or using external resources. Sentences are selected based on these topic labels in order to form a summary. Two selection methods are proposed in the paper. Experiments on DUC2004 dataset has shown that Vector-based methods are better, i.e. map topic labels and sentences to a word vector and a letter trigram vector space to find those sentences which are syntactically and semantically related with the topic labels in order to form a summary. Experiments show that the produced summaries are informative, abstractive and better than the baseline method.

Implicit discourse relation recognition is an extremely challenging task, for it lacks of explicit connectives between two arguments. Currently, most methods to address this problem can be regarded as to solve it in two stages, the first is to extract features from two arguments separately, and the next is to apply those features to some standard classifier. However, during the first stage, those methods neglect the links between two arguments and thus are blind to find pair-specified clues at the very beginning. This paper therefore makes an attempt to model sentence with its targeted pair in mind. Concretely, an LSTM model with attention mechanism is adapted to accomplish this idea. Experiments on the benchmark dataset show that without the help of feature engineering or any external linguistic knowledge, our proposed model outperforms previous state-of-the-art systems.