Growing parts of the population suffer from mental health problems and psychologists lack capacity to diagnose, let alone treat, all those in need of it. Given recent advancements in the field, deep learning-based NLP techniques could help by detecting those in need of help based on their written text. To this end, this work improves the current state-of-the-art Hierarchical Attention Network (HAN) model by incorporating contextual awareness through BERT-based word embeddings and a multi-head self-attention user-encoder yielding the Context-HAN model. When trained and tested on the eRisk data sets on Self-Harm, Anorexia, and Depression, Context-HAN outperformed the HAN model across all data sets based on various evaluation measures. Furthermore, we find and discuss some interesting insights from analysis of the attention scores, such as that longer and more recently written posts are more important for classification. This work shows the potential of attention mechanisms to leverage contextual information to improve the effectiveness of NLP methods at detecting mental health disorders from user-written text.

The Semantic Web aims to make information intelligible for computers. In the Semantic Web, unstructured information from text is represented using ontologies, such that computers can understand text better. However, adding text information to existing ontologies by hand is time-consuming. Information extraction rules can help to automate this process. In the process of learning information extraction rules, patterns are constructed that consist of lexico-syntactic and lexico-semantic features from text, which aim to extract Resource Description Framework subject-predicate-object expressions. In this paper, we investigate the following four metaheuristics for learning ontology-based information extraction rules: Particle Swarm Optimization, 2-Phase Optimization, Ant Colony Optimization, and Genetic Algorithm (GA). We evaluate all methods using financial news data. GA gives the best F1-measure results, but the other metaheuristics are faster.

Currently, deep learning models are commonly used for Aspect-Based Sentiment Analysis (ABSA). These deep learning models are often seen as black boxes, meaning that they are inherently difficult to interpret. To improve deep learning models, it is crucial to understand their inner workings. We aim to interpret black box models by implementing model-agnostic local interpretation methods. Inspired by Local Interpretable Model-agnostic Explanations (LIME) and Local Rule-based Explanations (LORE) and combined with a Similarity-based Sampling (SS) method, we propose SS-LIME and SS-LORE, and use Anchor to explain two state-of-the-art ABSA deep learning models. The deep learning models build upon the Left-Center-Right separated neural network with Rotatory attention (LCR-Rot) model, extended by iterating multiple times over the rotatory attention mechanism (LCR-Rot-hop) and hierarchical attention and context-dependent word embeddings (LCR-Rot-hop++). We evaluate the proposed models in terms of fidelity, hit rate, and user interpretability using the SemEval 2016 dataset consisting of restaurant reviews for ternary sentiment classification. Results show that the LCR-Rot-hop and LCR-Rot-hop++ models are best explained by SS-LIME and SS-LORE, respectively. Furthermore, we conclude that the LCR-Rot-hop++ model can be better interpreted than the LCR-Rot-hop model. • We explain the predictions of two deep learning models for sentiment analysis. • We provide interpretation models for ternary classification. • We use post-hoc classifiers with a homogeneous sampling method. • We extend the Submodular Pick algorithm by considering local importance.

Text summarization is an increasingly crucial focus of Natural Language Processing (NLP), and state-of-the-art models such as PEGASUS have demonstrated remarkable potential to ever more efficient and accurate abstractive summarization . Nonetheless, recent developments of deep learning models that focus on training with large datasets can become at risk of sub-optimal generalization, inefficient training time, and can get stuck at local optima due to high-dimensional non-convex optimization domains. Current research in the field of NLP suggests that leveraging curriculum learning techniques to guide model training (enabling the model to learn from training data with increasing difficulty) could provide a means to achieve enhanced model performance. In this paper we investigate the effectiveness of curriculum learning strategies and data augmentation techniques on PEGASUS to increase performance with low-resource training data from the CNN/DM dataset. We introduce a novel text-summary pair complexity scoring algorithm along with two simple baseline difficulty measures. We find that our novel complexity sorting method consistently outperforms the baseline sorting methods and boosts performance of PEGASUS. The Baby-Steps curriculum learning strategy with this sorting method leads to performance improvements of 5.65 %, from a combined ROUGE F1-score of 83.28 to 87.99. When this strategy is combined with a data augmentation technique, Easy Data Augmentation, this leads to an improvement to 6.54 %. These statistics are relative to a baseline without curriculum learning or data augmentation.

E-commerce has grown a lot in recent years and so has the research performed in this field. In this paper we estimate order satisfaction by predicting outcomes of relevant variables at the moment an order is made, such that companies can act on this signal. In order to deal with data that is not known at the order date (i.e., lagged missing data), we propose an extension of an existing generative imputation method. The Generative Adversarial Imputation Network (GAIN) is suitable for data imputation on tabular datasets. A more stable method is the Wasserstein GAIN (WGAIN). In this paper, we propose to improve this method by adding the Gradient Penalty to WGAIN resulting in WGAIN-GP. We perform experiments on a large dataset from a Dutch online retailer. Using WGAIN-GP we obtain a better accuracy of 61 % at the order date compared to 54 % and 53 % obtained by GAIN and WGAIN, respectively.

Sentiment analysis extracts the sentiment of content creators, enabling users to easily gain valuable insights from such data. Most existing methods rely on supervised learning approaches using labeled data. However, the retrieval of such labeled training data is difficult and expensive, especially for new domains and/or languages. This work focuses on simultaneously detecting aspect categories and sentiment polarities for a given sentence in a weakly-supervised setting. Two methods are proposed that combine an unsupervised labeling algorithm with a neural network architecture. The first proposed two-step model (SB-ASC) takes seed sentences as input for the labeling algorithm. By leveraging the power of pre-trained Sentence-BERT embeddings, the method is able to understand the contextual meaning of sentences to create a high-quality labeled dataset. This dataset is used by a class imbalance-robust BERT-based neural network that jointly learns latent features of aspect categories and the corresponding sentiment. The second proposed method (WB-ASC) uses the same neural network structure but takes seed words instead of seed sentences as input for the labeling algorithm. We conclude that SB-ASC outperforms WB-ASC as well as baselines and state-of-the-art weakly-supervised methods for aspect sentiment detection, achieving F1 scores for aspect category detection of 71.35%, 86.99%, and 73.86%, and F1 scores for sentiment classification of 89.24%, 89.98%, and 75.58% for the SemEval 2016 restaurant-5, restaurant-3, and laptop datasets, respectively. Furthermore, using domain-specific contextual language models boosts performance. • We predict aspect categories and sentiment polarities using weak supervision. • We propose two weak supervision methods. • We find that using seed sentences instead of words as input improves predictions. • We show that using domain-specific contextual language models boosts performance.

Automated feature engineering (AutoFE) and automated machine learning (AutoML) can both be used in a machine learning project to improve the efficiency of a data scientist. In recent years, different algorithms have been developed for both sub fields independently of each other. In this study, the use of AutoFE in combination with AutoML has been evaluated for the first time to determine if AutoFE can increase the model accuracy, while not increasing the computation time. A data fusion meta-learning approach was extended, generalized, and applied to an AutoFE method and then further combined with a pre-existing AutoML method. In the meta-learning approach, more than 150 online data sets were used to create models that recommended the best operator to apply to a certain feature. Using twelve evaluation data sets, we show that combining AutoFE and AutoML is indeed valuable. The accuracy measure used was increased on average by 0.54% compared to using AutoML alone. For multiple data sets, the use of AutoFE significantly outperformed a strategy in which no feature engineering was done, while in the remaining data sets it never significantly performed worse. Therefore, it can be concluded that it is beneficial to combine this computationally efficient AutoFE method with AutoML.

Abstract As databases become larger and less structured, the JavaScript Object Notation (JSON) data format has risen in usage compared to other data formats like XML. At the same time, while extracting data from these large datasets efficiently is of obvious importance, there has been far less research regarding the optimization of JSON queries than there has relating to the querying of XML data. Thus a JSON Data Model and JSON Algebra (JAL) are proposed, as well as a heuristic optimization algorithm, for the purpose of improving the efficiency of queries of JSON data. We implement the proposed algorithm and compare the efficiency gain that it provides in terms of both the theoretical and physical cost of executing queries. We find that the algorithm significantly reduces query costs compared to an unoptimized baseline. Additionally, we find that the efficiency gain is considerably larger when querying databases with many documents than those with relatively fewer documents.

One of the problems that e-commerce users face is that the desired products are sometimes not available and Web shops fail to provide similar products due to their exclusive reliance on Boolean faceted search. User preferences are also often not taken into account. In order to address these problems, we present a novel framework specifically geared towards approximate faceted search within the product catalog of a Web shop. It is based on adaptations to the p-norm extended Boolean model, to account for the domain-specific characteristics of faceted search in an e-commerce environment. These e-commerce specific characteristics are, for example, the use of quantitative properties and the presence of user preferences. Our approach explores the concept of facet similarity functions in order to better match products to queries. In addition, the user preferences are used to assign importance weights to the query terms. Using a large-scale experimental setup based on real-world data, we conclude that the proposed algorithm outperforms the considered benchmark algorithms. Last, we have performed a user-based study in which we found that users who use our approach find more relevant products with less effort.

This work researches the drivers of e-commerce conversion of one of the largest e-commerce companies in the Netherlands. We focus on product page conversion, i.e., the probability that a customer who visits a specific product page also buys the product. This probability differs between products, which we explain by a variety of factors like pricing, delivery times, reviews, seller type, and the quality of the content. Understanding the drivers of conversion is the first step in increasing it, and therefore an important step in increasing overall sales. We describe the process of transforming Big Data into valuable insights using Bayesian statistics and apply a Bayesian Binomial model to a dataset of 15 million records using a distributed MCMC algorithm. In addition, we apply a simple Binomial GLM on a small sample of these 15 million records for comparison. We find that using the full 15 million records results in a significant reduction in the variance of the estimated parameters. In terms of the actual drivers of e-commerce conversion, we observe that products with competitive pricing, short delivery times, and good content all achieve a higher conversion on average. Furthermore, reviews, average review score, and the number of reviews for a product have a large positive effect on conversion compared to other characteristics. This provides a unique insight into what makes people decide whether or not to purchase a product.

Recommender systems are widely used in webshops because of their ability to provide users with personalized recommendations. However, the cold-user problem (i.e., recommending items to new users) is an important issue many webshops face. With the recent General Data Protection Regulation in Europe, the use of additional user information such as demographics is not possible without the user’s explicit consent. Several techniques have been proposed to solve the cold-user problem. Many of these techniques utilize Active Learning (AL) methods, which let cold users rate items to provide better recommendations for them. In this research, we propose two novel approaches that combine reinforcement learning with AL to elicit the users’ preferences and provide them with personalized recommendations. We compare reinforcement learning approaches that are either AL-based or item-based, where the latter predicts users’ ratings of an item by using their ratings of similar items. Differently than many of the existing approaches, this comparison is made based on implicit user information. Using a large real-world dataset, we show that the item-based strategy is more accurate than the AL-based strategy as well as several existing AL strategies.

Every day millions of news articles and (micro)blogs that contain financial information are posted online. These documents often include insightful financial aspects with associated sentiments. In this paper, we predict financial aspect classes and their corresponding polarities (sentiment) within sentences. We use data from the Financial Question & Answering (FiQA) challenge, more precisely the aspect-based financial sentiment analysis task. We incorporate the hierarchical structure of the data by using the parent aspect class predictions to improve the child aspect class prediction (two-step model). Furthermore, we incorporate model output from the child aspect class prediction when predicting the polarity. We improve the F1 score by 7.6% using the two-step model for aspect classification over direct aspect classification in the test set. Furthermore, we improve the state-of-the-art test F1 score of the original aspect classification challenge from 0.46 to 0.70. The model that incorporates output from the child aspect classification performs up to par in polarity classification with our plain RoBERTa model. In addition, our plain RoBERTa model outperforms all the state-of-the-art models, lowering the MSE score by at least 28% and 33% for the cross-validation set and the test set, respectively.

E-commerce companies collaborate with retailers to sell products via their platforms, making it increasingly important to preserve platform quality. In this paper, we contribute by introducing a novel method to predict the quality of product orders shortly after they are placed. By doing so, platforms can act fast to resolve bad quality orders and potentially prevent them from happening. This introduces a trade-off between accuracy and timeliness, as the sooner we predict, the less we know about the status of a product order and, hence, the lower the reliability. To deal with this, we introduce the Hierarchical Classification Over Time (HCOT) algorithm, which dynamically classifies product orders using top-down, non-mandatory leaf-node prediction. We enforce a blocking approach by proposing the Certainty-based Automated Thresholds (CAT) algorithm, which automatically computes optimal thresholds at each node. The resulting CAT-HCOT algorithm has the ability to provide both accurate and timely predictions by classifying a product order's quality on a daily basis if the classification reaches a predefined certainty. CAT-HCOT obtains a predictive accuracy of 94%. Furthermore, CAT-HCOT classifies 40% of product orders on the order date itself, 80% within five days after the order date, and 100% of product orders after 10 days.

The current models proposed for aspect-based sentiment classification (ABSC) are mainly developed with the purpose of providing high rates of accuracy, regardless of the inner working which is usually difficult to understand. Considering the state-of-art model LCR-Rot-hop++ for ABSC, we use diagnostic classifiers to gain insights into the encoded information of each layer. Starting from a set of various hypotheses, we test how sentiment-related information is captured by different layers of the model. Given the model architecture, information about the related words to the target is easily extracted. Also, the model is able to detect to some extent information about the sentiments of the words and, in particular, sentiments of the words related to the target. However, the model is less effective in extracting the aspect mentions associated with a word and the general structure of the sentence.

Attention is an important mechanism that can be employed for a variety of deep learning models across many different domains and tasks. This survey provides an overview of the most important attention mechanisms proposed in the literature. The various attention mechanisms are explained by means of a framework consisting of a general attention model, uniform notation, and a comprehensive taxonomy of attention mechanisms. Furthermore, the various measures for evaluating attention models are reviewed, and methods to characterize the structure of attention models based on the proposed framework are discussed. Last, future work in the field of attention models is considered.

Sentiment analysis techniques are becoming more and more important as the number of reviews on the World Wide Web keeps increasing. Aspect-based sentiment analysis (ABSA) entails the automatic analysis of sentiments at the highly fine-grained aspect level. One of the challenges of ABSA is to identify the correct sentiment expressed towards every aspect in a sentence. In this paper, a neural attention model is discussed and three extensions are proposed to this model. First, the strengths and weaknesses of the highly successful CABASC model are discussed, and three shortcomings are identified: the aspect-representation is poor, the current attention mechanism can be extended for dealing with polysemy in natural language, and the design of the aspect-specific sentence representation is upheld by a weak construction. We propose the Extended CABASC (E-CABASC) model, which aims to solve all three of these problems. The model incorporates a context-aware aspect representation, a multi-dimensional attention mechanism, and an aspect-specific sentence representation. The main contribution of this work is that it is shown that attention models can be improved upon using some relatively simple extensions, such as fusion gates and multi-dimensional attention, which can be implemented in many state-of-the-art models. Additionally, an analysis of the parameters and attention weights is provided.

In the current literature, many methods have been devised for sentiment quantification. In this work, we propose AspEntQuaNet, one of the first methods for aspect-based sentiment quantification. It extends the state-of-the-art QuaNet deep learning method for sentiment quantification in two ways. First, it considers aspects and ternary sentiment quantification concerning these aspects instead of binary sentiment quantification. Second, it improves on the results of QuaNet with an entropy-based sorting procedure instead of multisorting. Other sentiment quantification methods have also been adapted for ternary sentiment quantification instead of binary sentiment quantification. Using the modified version of the SemEval 2016 dataset for aspect-based sentiment quantification, we show that AspEntQuaNet is superior to all other considered existing methods based on obtained results for various aspect categories. In particular, AspEntQuaNet outperforms QuaNet often by a factor of 2 on all considered evaluation measures.

Due to low switching costs and stiff competition, customer relationship management has become a central component in the marketing strategy of telecommunication service providers. Since the costs of acquiring a new customer are five times higher than the costs of maintaining an existing customer, telecommunication service providers are eager to reduce the churn rate. A solid understanding of customer churn behavior can help to address this problem. Reducing the churn rate can translate into significant revenue gains and might provide the edge to outperform the competitor. In this paper, we predict the propensity to churn for customers of a Dutch telecommunication service provider by employing a duration model. While predicting churn, we simultaneously predict the reason for which the customer churns, using a competing risks model. Since the telecommunication service provider has valuable textual data based on transcripts of calls between customers and the customer service center, we incorporate topics extracted from this textual data as variables in our models, by employing Latent Dirichlet Allocation (LDA). We compare four models and find that the models that have incorporated topic variables usually yield the best churn forecasts. Also, the investigated models beat the considered benchmark model, which is the model currently deployed at the telecommunication service provider.

State-of-the-art decision support systems for stock price prediction incorporate pattern-based event detection in text into their predictions. These systems typically fail to account for word meaning, even though word sense disambiguation is crucial for text understanding. Therefore, we propose an advanced natural language processing pipeline for event-based stock price prediction, that allows for word sense disambiguation to be incorporated in the event detection process. We identify events in natural language news messages and subsequently weight these events for their historical impact on stock prices. We assess the merit of word sense disambiguation in event-based stock price prediction in two evaluation scenarios for NASDAQ-100 companies, based on historical stock prices and news articles retrieved from Dow Jones Newswires over a 2-year period. We evaluate the precision of generated buy and sell signals based on our predicted stock price movements, as well as the excess returns generated by a trading strategy that acts upon these signals. Event-based stock price predictions seem most reliable about 2 days into the future. The number of detected events tends to reduce with over 30% when graph-based word sense disambiguation using a degree centrality measure is applied in the event detection process, thus reducing the noise introduced into the stock price movement predictions by high-impact ambiguous events. As a result, modest improvements in the precision of buy and sell signals generated based on these predictions tend to lead to vast improvements of on average about 70% in the associated excess returns.

To keep up with the growing interest of using Web shops for product comparison, we have developed a method that targets the problem of product duplicate detection. If duplicates can be discovered correctly and quickly, customers can compare products in an efficient manner. We build upon the state-of-the-art Multi-component Similarity Method (MSM) for product duplicate detection by developing an automated pre-processing phase that occurs before the similarities between products are calculated. Specifically, in this prior phase the features of products are aligned between Web shops, using metrics such as the data type, coverage, and diversity of each key, as well as the distribution and used measurement units of their corresponding values. With this information, the values of these keys can be more meaningfully and efficiently employed in the process of comparing products. Applying our method to a real-world dataset of 1629 TV’s across 4 Web shops, we find that we increase the speed of the product similarity phase by roughly a factor 3 due to fewer meaningless comparisons, an improved brand analyzer, and a renewed title analyzer. Moreover, in terms of quality of duplicate detection, we significantly outperform MSM with an F1-measure of 0.746 versus 0.525.

Aspect-based sentiment analysis allows one to compute the sentiment for an aspect in a certain context. One problem in this analysis is that words possibly carry different sentiments for different aspects. Moreover, an aspect’s sentiment might be highly influenced by the domain-specific knowledge. In order to tackle these issues, in this paper, we propose a hybrid solution for sentence-level aspect-based sentiment analysis using A Lexicalized Domain Ontology and a Regularized Neural Attention model (ALDONAr). The bidirectional context attention mechanism is introduced to measure the influence of each word in a given sentence on an aspect’s sentiment value. The classification module is designed to handle the complex structure of a sentence. The manually created lexicalized domain ontology is integrated to utilize the field-specific knowledge. Compared to the existing ALDONA model, ALDONAr uses BERT word embeddings, regularization, the Adam optimizer, and different model initialization. Moreover, its classification module is enhanced with two 1D CNN layers providing superior results on standard datasets.

Customers of a webshop are often presented large assortments, which can lead to customers struggling finding their desired product(s), an issue known as choice overload. In order to overcome this issue, recommender systems are used in webshops to provide personalized product recommendations to customers. Though, model-based recommender systems are not able to provide recommendations to new customers (i.e., cold users). To facilitate recommendations to cold users we investigate multiple active learning strategies, and subsequently evaluate which active learning strategy is able to optimally elicit the preferences from the cold users in a matrix factorization context. Our model is empirically validated using a dataset from the webshop of de Bijenkorf, a Dutch department store. We find that the overall best-performing active learning strategy is PopError, an active learning strategy that measures the variance score for each item.

This work focuses on sentence-level aspect-based sentiment analysis for restaurant reviews. A two-stage sentiment analysis algorithm is proposed. In this method, first a lexicalized domain ontology is used to predict the sentiment and as a back-up algorithm a neural network with a rotatory attention mechanism (LCR-Rot) is utilized. Furthermore, two features are added to the backup algorithm. The first extension changes the order in which the rotatory attention mechanism operates (LCR-Rot-inv). The second extension runs over the rotatory attention mechanism for multiple iterations (LCR-Rot-hop). Using the SemEval-2015 and SemEval-2016 data, we conclude that the two-stage method outperforms the baseline methods, albeit with a small percentage. Moreover, we find that the method where we iterate multiple times over a rotatory attention mechanism has the best performance.

The ever-increasing amount of Web information offered to news readers (e.g., news analysts) stimulates the need for news selection, so that informed decisions can be made with up-to-date knowledge.Hermes is an ontologybased framework for building news personalization services.It uses an ontology crafted from available news sources, allowing users to select and filter interesting concepts from a domain ontology.The Aethalides framework enhances the Hermes framework by enabling news classification through lexicographic and semantic properties.For this, Aethalides applies word sense disambiguation and ontology learning methods to news items.When tested on a set of news items on finance and politics, the Aethalides implementation yields a precision and recall of 74.4% and 49.4%, respectively, yielding an F 0.5 -measure of 67.6% when valuing precision more than recall.

Currently many webshops rely on a fixed list of product facets to help users find the products of interest. Such a solution suffers from two problems: (1) it is difficult to devise a fixed list of facets that would satisfy all user interests, and (2) the top facets could become obsolete when all product results have these facets. To address these two problems we propose a novel algorithm for dynamic ordering of the product facets based on the query results. This algorithm relies on measures such as specificity and dispersion for qualitative and quantitative facets, respectively, to rank the properties associated with these facets so that users are able to find the products of interest with a minimum number of drill-down steps. Using a large-scale simulation study and a user-based evaluation, we show that our algorithm outperforms the expert-based fixed facets approach, a greedy baseline, and a state-of-the-art entropy-based solution. This paper is an extended abstract of our previous work [1].

Using online consumer reviews as electronic word of mouth to assist purchase-decision making has become increasingly popular. The Web provides an extensive source of consumer reviews, but one can hardly read all reviews to obtain a fair evaluation of a product or service. A text processing framework that can summarize reviews, would therefore be desirable. A subtask to be performed by such a framework would be to find the general aspect categories addressed in review sentences, for which this paper presents two methods. In contrast to most existing approaches, the first method presented is an unsupervised method that applies association rule mining on co-occurrence frequency data obtained from a corpus to find these aspect categories. While not on par with state-of-the-art supervised methods, the proposed unsupervised method performs better than several simple baselines, a similar but supervised method, and a supervised baseline, with an -score of 67%. The second method is a supervised variant that outperforms existing methods with an -score of 84%.

Customers of a webshop are often presented large assortments, which can lead to customers struggling finding their desired product(s), an issue known as choice overload. In order to overcome this issue, recommender systems are used in webshops to provide personalized product recommendations to customers. Though, recommender systems using matrix factorization are not able to provide recommendations to new customers (i.e., cold users). To facilitate recommendations to cold users we investigate multiple active learning strategies, and subsequently evaluate which active learning strategy is able to optimally elicit the preferences from the cold users. Our model is empirically validated using a dataset from the webshop of de Bijenkorf, a Dutch department store. We find that the overall best-performing active learning strategy is PopGini, an active learning strategy which combines the popularity of an item with its Gini impurity score.

Faceted browsing is widely used in Web shops and product comparison sites. In these cases, a fixed ordered list of facets is often employed. This approach suffers from two main issues. First, one needs to invest a significant amount of time to devise an effective list. Second, with a fixed list of facets, it can happen that a facet becomes useless if all products that match the query are associated to that particular facet. In this work, we present a framework for dynamic facet ordering in e-commerce. Based on measures for specificity and dispersion of facet values, the fully automated algorithm ranks those properties and facets on top that lead to a quick drill-down for any possible target product. In contrast to existing solutions, the framework addresses e-commerce specific aspects, such as the possibility of multiple clicks, the grouping of facets by their corresponding properties, and the abundance of numeric facets. In a large-scale simulation and user study, our approach was, in general, favorably compared to a facet list created by domain experts, a greedy approach as baseline, and a state-of-the-art entropy-based solution.

The field of sentiment analysis, in which sentiment is gathered, analyzed, and aggregated from text, has seen a lot of attention in the last few years. The corresponding growth of the field has resulted in the emergence of various subareas, each addressing a different level of analysis or research question. This survey focuses on aspect-level sentiment analysis, where the goal is to find and aggregate sentiment on entities mentioned within documents or aspects of them. An in-depth overview of the current state-of-the-art is given, showing the tremendous progress that has already been made in finding both the target, which can be an entity as such, or some aspect of it, and the corresponding sentiment. Aspect-level sentiment analysis yields very fine-grained sentiment information which can be useful for applications in various domains. Current solutions are categorized based on whether they provide a method for aspect detection, sentiment analysis, or both. Furthermore, a breakdown based on the type of algorithm used is provided. For each discussed study, the reported performance is included. To facilitate the quantitative evaluation of the various proposed methods, a call is made for the standardization of the evaluation methodology that includes the use of shared data sets. Semanticallyrich concept-centric aspect-level sentiment analysis is discussed and identified as one of the most promising future research direction.

We present a general approach for Web news items analysis in relation to stock prices. The framework that we introduce provides the ability to study the impact of events extracted from news on stock prices. The relation between events and price is quantified in terms of the i) paired-samples t-test, ii) McNemar's test, and iii) confidence and support. The extraction, representation, and visualization of data are key components of the proposed framework. The validation of the framework is based on three case studies, involving Tesco, Shell, and British Petroleum, and the price reaction(s) to different news events.

In this paper we present a framework for automatic exploitation of news in stock trading strategies. Events are extracted from news messages presented in free text without annotations. We test the introduced framework by deriving trading strategies based on technical indicators and impacts of the extracted events. The strategies take the form of rules that combine technical trading indicators with a news variable, and are revealed through the use of genetic programming. We find that the news variable is often included in the optimal trading rules, indicating the added value of news for predictive purposes and validating our proposed framework for automatically incorporating news in stock trading strategies.

Multifaceted search is a commonly used interaction paradigm in e-commerce applications, such as Web shops. Because of the large amount of possible product attributes, Web shops usually make use of static information to determine which facets should be displayed. Unfortunately, this approach does not take into account the user query, leading to a non-optimal facet drill down process. In this paper, we focus on automatic facet selection, with the goal of minimizing the number of steps needed to find the desired product. We propose several algorithms for facet selection, which we evaluate against the state-of-the-art algorithms from the literature. We implement our approach in a Web application called faccy.net. The evaluation is based on simulations employing 1000 queries, 980 products, 487 facets, and three drill down strategies. As evaluation metrics we use the average number of clicks, the average utility, and the top-10 promotion percentage. The results show that the Probabilistic Entropy algorithm significantly outperforms the other considered algorithms.

As today’s financial markets are sensitive to breaking news on economic events, accurate and timely automatic identification of events in news items is crucial. Unstructured news items originating from many heterogeneous sources have to be mined in order to extract knowledge useful for guiding decision making processes. Hence, we propose the Semantics-Based Pipeline for Economic Event Detection (SPEED), focusing on extracting financial events from news articles and annotating these with meta-data at a speed that enables real-time use. In our implementation, we use some components of an existing framework as well as new components, e.g., a high-performance Ontology Gazetteer, a Word Group Look-Up component, a Word Sense Disambiguator, and components for detecting economic events. Through their interaction with a domain-specific ontology, our novel, semantically enabled components constitute a feedback loop which fosters future reuse of acquired knowledge in the event detection process.

SUMMARY In this paper we propose the Automatic Review Recognition and annOtation of Webpages method, which is capable of identifying and annotating user submitted reviews on a Web page. The method consists of six steps: data preparation, page‐level pattern identification, subjectivity analysis, container structure analysis, review properties identification, and review annotation. For the evaluation, we have implemented the method and tested it on various review websites. On the basis of the performed evaluation, we conclude that our method is capable of identifying and annotating the majority of reviews. Copyright © 2012 John Wiley & Sons, Ltd.

Value-at-Risk (VaR) is an important tool to assess portfolio risk. When calculating VaR based on historical stock return data, we hypothesize that this historical data is sensitive to outliers caused by news events in the sampled period. In this paper, we research whether the VaR accuracy can be improved by considering news events as additional input in the calculation. This involves processing the historical data in order to reflect the impact of news on the stock returns. Our experiments show that when an event occurs, removing the noise (that is caused by an event) from the measured stock prices for a small time window can improve VaR predictions.

Building domain taxonomies is a crucial task in the domain of ontology construction. Domain taxonomy learning keeps getting more important as a form of automatically obtaining a knowledge representation of a certain domain. The alternative of manually developing domain taxonomies is not trivial. The main issues encountered when manually developing a taxonomy are the non-availability of a domain knowledge expert and the considerable amount of effort needed for this task. This paper proposes Taxo Learn, an approach to automatic construction of domain taxonomies. Taxo Learn is a new methodology that combines aspects from existing approaches, but also contains new steps in order to improve the quality of the resulted domain taxonomy. The contribution of this paper is threefold. First, we employ a word sense disambiguation step when detecting concepts in the text. Second, we show the use of semantics-based hierarchical clustering for the purpose of taxonomy learning. Third, we propose a novel dynamic labeling procedure for the concept clusters. We evaluate our approach by comparing the machine generated taxonomy with a manually constructed golden taxonomy. Based on a corpus of documents in the field of financial economics, Taxo Learn shows a high precision for the learned taxonomic concept relationships.

In order to effectively and efficiently disclose the ever-growing amount of widely distributed RDF data to demanding users in real-time environments, RDF query engines need to optimize the join order of partial query results. For this, a two-phase optimization (2PO) algorithm and a genetic algorithm (GA) have already been proposed. We propose an alternative approach - an ant colony system (ACS). On a large RDF data source, our approach significantly outperforms both 2PO and the GA in terms of execution time and solution quality for RDF chain queries consisting of up to about ten joins. For larger queries, our novel ACS delivers solutions of better quality than 2PO does, while realizing a solution quality that is comparable to the solution quality of the GA method. However, the GA approach offers the best trade-off between execution time and solution quality for such larger queries.

In this paper we evaluate several novelty control mechanisms for ranking Web news articles depicting a story. These mechanisms rank individual news items based on a novelty measure using as context the items which have been previously reviewed. An evaluation within the Hermes news personalization framework is performed for pair wise and non-pair wise novelty control mechanisms based on various distance measures and vector-based news representations. On average, the most effective distance measure is Kullback-Leibler and the best performing news representation vector uses named entities.

Abstract Through its interoperability and reasoning capabilities, the Semantic Web opens a realm of possibilities for developing intelligent systems on the Web. The Web Ontology Language (OWL) is the most expressive standard language for modeling ontologies, the cornerstone of the Semantic Web. However, up until now, no standard way of expressing time and time-dependent information in OWL has been provided. In this paper, we present a temporal extension of the very expressive fragment ${cal SHIN}({cal D})$ of the OWL Description Logic language, resulting in the temporal OWL language. Through a layered approach, we introduce three extensions: 1) concrete domains, which allow the representation of restrictions using concrete domain binary predicates; 2) temporal representation, which introduces time points, relations between time points, intervals, and Allen's 13 interval relations into the language; and 3) timeslices/fluents, which implement a perdurantist view on individuals and allow for the representation of complex temporal aspects, such as process state transitions. We illustrate the expressiveness of the newly introduced language by using an example from the financial domain.

In this paper we build on previous work related to predicting the MSCI EURO index based on content analysis of ECB statements. Our focus is on reducing the number of features employed for prediction through feature selection. For this purpose we rely on two methodologies: (stepwise) linear regression and greedy forward feature subset selection. The original dataset consists of 13 features (General Inquirer content categories). Both methodologies provide an improvement in the overall accuracy of the model, while reducing the number of features employed. Through linear regression we achieve an accuracy of 67.58% on the testing set by relying on six features, while greedy forward selection enables an accuracy on the test set of 69.50% while relying on eight features.

Sentiment analysis has applications in many areas and the exploration of its potential has only just begun. We propose Pathos, a framework which performs document sentiment analysis (partly) based on a document's discourse structure. We hypothesize that by splitting a text into important and less important text spans, and by subsequently making use of this information by weighting the sentiment conveyed by distinct text spans in accordance with their importance, we can improve the performance of a sentiment classifier. A document's discourse structure is obtained by applying Rhetorical Structure Theory on sentence level. When controlling for each considered method's structural bias towards positive classifications, weights optimized by a genetic algorithm yield an improvement in sentiment classification accuracy and macro-level F1 score on documents of 4.5% and 4.7%, respectively, in comparison to a baseline not taking into account discourse structure.

A key element for decision makers to track is their stakeholders' sentiment. Recent developments show a tendency of including various aspects other than word frequencies in automated sentiment analysis approaches. One of these aspects is negation, which can be accounted for in various ways. We compare several approaches to accounting for negation in sentiment analysis, differing in their methods of determining the scope of influence of a negation keyword. On a set of English movie review sentences, the best approach is to consider two words, following a negation keyword, to be negated by that keyword. This method yields a significant increase in overall sentiment classification accuracy and macro-level F1of 5.5% and 6.2%, respectively, compared to not accounting for negation. Additionally optimizing sentiment modification of negated words to a value of -1.27 rather than -1 yields a significant 7.1% increase in accuracy and a significant 8.0% increase in macro-level F1.

In this paper we investigate whether the MSCI EURO index can be predicted based on the content of European Central Bank (ECB) statements. We propose a new model to retrieve information from free text and transform it into a quantitative output. For this purpose, we first identify all adjectives in an ECB statement by using the Stanford Part-of-Speech Tagger and feed these to the General Inquirer (GI) content analysis tool. From GI we obtain a matrix that provides for each document and for each content category the percentage of words in the document that fall under each category. After normalizing the data, we develop a Takagi-Sugeno (TS) fuzzy model using fuzzy c-means clustering. The TS fuzzy system is used to model the levels of the MSCI EURO index. To determine the performance of the model, we focus on the accuracy of predicting upward or downward movement in the index, and obtain, on average, an accuracy of 66%, that corresponds to an improvement of 16% over a random classifier.

The emergence of Web 2.0 and the semantic Web as established technologies is fostering a whole new breed of Web applications and systems. These are often centered around knowledge engineering and context awareness. However, adequate temporal formalisms underlying context awareness are currently scarce. Our focus in this paper is two-fold. We first introduce a new OWL-based temporal formalism - TOWL - for the representation of time, change, and state transitions. Based hereon we present a financial Web-based application centered around the aggregation of stock recommendations and financial data.

One of the steps in the knowledge domain visualization process is the display of the low-dimensional space in which the items under analysis are positioned. In this paper we investigate the visualization of concept associations by means of concept clustering and edge bundling, two techniques that help alleviate the cluttering caused by concept labels and edges. Hierarchical clustering is used for the identification of the different domain categories (i.e., clusters). The edges between pairs of concepts belonging to two clusters are aggregated in a single edge connecting the clusters’ centers. The proposed approach enables us to discover the hierarchical structure of our knowledge domain and analyze the strength between the different domain categories.

The World Wide Web is a rapidly growing business environmenthosting a large number of business transactions. Methods fordesigning Web Information Systems (WIS) have adopted processmodels, typically as extensions of the navigation models they arebased on. We observe that the structure of business processes inWIS goes beyond the scope of the navigation structure and deservesto be a more prominent aspect of the design, abstracted from thenavigation specification. This paper explains the design of WISdriven by the specification of business processes of WIS that isused for the automatic generation of models describing theapplication logic including the navigation structure. Thespecification of processes is explained on a conference review andsubmission system and the generation process is demonstrated onthe Hera application model.

The concept mapping algorithm proposed in an earlier paper is one of the dimensionality reduction techniques that can be used for knowledge domain visualization. Using this algorithm to visualize large knowledge domains may not always provide a good overview of the domain due to visual cluttering of concepts. In this paper, we propose to apply kernel density estimation to the visualization of concept maps in order to be able to better explore large knowledge domains. Kernel density estimation proves to be useful for the identification of concept clusters at different levels of detail. In addition to the visual exploration of large knowledge domains, we are also able to visually verify the hypothesis that the concept mapping algorithm places related concepts close to each other. The flexibility and effectiveness of our approach is validated by applying the proposed technique to different visualization scenarios for the field of computational intelligence

As Resource Description Framework (RDF) reaches maturity, there is an increasing need for tools that support it. A common and natural representation for RDF data is a directed labeled graph. Although there are tools to edit and/or browse RDF graph representations, we found their architecture rigid and not easily amenable to producing effective visual representations, especially for large RDF graphs. We discuss here how GViz, a general purpose graph visualisation tool, allows the easy construction and fine-tuning of various visual exploratory scenarios for RDF data. GViz's extended ability of customizing the visualisation's icons showed to be very useful in the context of RDF graph structures visualisation. Among the presented applications, we mention customizable selections, schema-instance comparison, instances comparison, and schemas comparison (schema evolution). GViz proved to be able not only to visualize large RDF data models, but also to be very flexible in designing scenario-specific queries to support the exploration process.

To make the WWW machine-understandable there is a strong demand both for languages describing metadata and for languages querying metadata. The Resource Description Framework (RDF), a language proposed by W3C, can be used for describing metadata about (Web) resources. The RDF schema (RDFS) extends RDF by providing means for creating application specific vocabularies (ontologies). While the above two languages are widely acknowledged as a standard means for describing Web metadata, a standardized language for querying RDF metadata is still an open issue. Research groups from both industry and academia are presently involved in proposing several RDF query languages. Due to the lack of an RDF algebra such query languages use APIs to describe their semantics and optimization issues are mostly neglected. This paper proposes RAL, an RDF algebra suitable for defining (and comparing) the semantics of different RDF query languages and (at a future stage) for performing algebraic optimizations. After a definition of the data model we present the operators with which the model can be manipulated. The operators come in three flavors: extraction operators retrieve needed resources from the input RDF model, loop operators support repetition, and construction operators build the resulting RDF model.