This paper deals with extracting and learning patterns of assonance as a stylistic feature for author attribution of historical texts. We describe an assonance extraction algorithm, and consider results from an extensive set of machine learning experiments, based on a historical corpus of 18th century American and British texts. The results are compared with those obtained from the use of other prosodic and traditional stylistic features.

Semantic similarity is a major automated approach to address many tasks such as essay grading, answer assessment, text summarization and information retrieval. Many semantic similarity methods rely on semantic representation such as Latent Semantic Analysis (LSA), an unsupervised method to infer a vectorial semantic representation of words or larger texts such as documents. Two ingredients in obtaining LSA vectorial representations are the corpus of texts from which the vectors are derived and the dimensionality of the resulting space. In this work, we investigate the effect of corpus size and vector dimensionality on assessing student generated content in advanced learning systems, namely, virtual internships. Automating the assessment of student generated content would greatly increase the scalability of virtual internships to millions of learners at reasonable costs. Prior work on automated assessment of notebook entries relied on classifiers trained on participant data. However, when new virtual internships are created for a new domain, for instance, no participant data is available a priori. Here, we report on our effort to develop a LSA-based assessment method without student data. Furthermore, we investigate the optimum corpus size and vector dimensionality for these LSA-based methods.

In this work we introduce the problem of incendiary news detection. We compare and contrast this problem with the problem of hate speech detection in social media. Most of the social media posts that are classified as hate speech contain straightforward slurs, insults, swearing, etc. In contrast to social media posts, incendiary news articles often do not contain any straightforward slurs and insults but, nevertheless, incite hate. To detect such news articles, we leverage are source where activists attempt to combat hate on-line by manually tagging the news articles inciting hate. We collect non-incendiary news by retrieving news articles from the websites of the news agencies which are recognized world-wide as serious media that are highly unlikely to contain foul language (BBC, CNN). We run a classification experiment using several classification approaches. We demonstrate that our system differentiates between incendiary and non-incendiary news with 97.0% accuracy. We ensure the validity of our approach by using two different non-incendiary news corpora.

In this paper we adapt the semi-supervised deep learning architecture known as Convolutional Ladder Networks, from the domain of computer vision, and explore how well it works for a semi-supervised Named Entity Recognition and Classification task with legal data. The idea of exploring a semi-supervised technique is to asses the impact of large amounts of unsupervised data (cheap to obtain) in specific tasks that have little annotated data, in order to develop robust models that are less prone to overfitting. In order to achieve this, first we must check the impact on a task that is easier to measure. We are presenting some preliminary results, however, the experiments carried out show some very interesting insights that foster further research in the topic.

Methods that calculate dense vector representations for text have proven to be very successful for knowledge representation. We study how to estimate dense representations for multi-modal data (e.g., text, continuous, categorical). We propose Feat2Vec as a novel model that supports supervised learning when explicit labels are available, and self-supervised learning when there are no labels. Feat2Vec calculates embeddings for data with multiple feature types, enforcing that all embeddings exist in a common space. We believe that we are the first to propose a method for learning self-supervised embeddings that leverage the structure of multiple feature types. Our experiments suggest that Feat2Vec outperforms previously published methods, and that it may be useful for avoiding the cold-start problem.

Despite a tremendous amount of work in the literature and in the commercial sectors, current approaches to multi-modal trip planning still fail to consistently generate plans that users deem optimal in practice. We believe that this is due to the fact that current planners fail to capture the true preferences of users, e.g., their preferences depend on aspects that are not modeled. An example of this could be a preference not to walk through an unsafe area at night. We present a novel multi-modal trip planner that allows users to up- load auxiliary geographic data (e.g., crime rates) and to specify temporal constraints and preferences over these data in combination with typical metrics such as time and cost. Concretely, our planner supports the modes walking, biking, driving, public transit, and taxi, uses linear temporal logic to capture temporal constraints, and preferential cost functions to represent preferences. We show by examples that this allows the expression of very interesting preferences and constraints that, naturally, lead to quite diverse optimal plans.

Learning strategies to address problems on graph and tree structures with no a-priori size limitations in cases where no known solution exists (and thus supervised data is hard to obtain), is a difficult problem with potential applications in a wide range of domains ranging from biological networks to protein folding and social network search. The main challenges here arise from the variable size representation that needs to be resolved in the context of Reinforcement Learning (RL) to address the problem. In this paper we consider a common, specific tree problem and show that it can be addressed using a combination of feature engineering and carefully designed learning processes. In particular, We consider the classical Nearest Neighbor Interchange (NNI) distance between unrooted labeled trees, which is defined as the minimum-cost sequence of operations that transform one tree into another. We introduce a representation and a reinforcement learning method that learns the transition dynamics and iteratively changes an arbitrary initial labeled tree into a goal configuration reachable through NNI. The differential tree representation and NNI actions permits the system to learn a strategy that is applicable to arbitrary sized trees. To train the system, we introduce a training process that uses randomly sampled trajectories to incrementally train more and more complex problems to overcome the difficulty of the overall strategy space. Experiments performed show that the system can successfully learn a strategy for effective NNI on complex trees.

Many software systems run on long-lifespan platforms that operate in diverse and dynamic environments. If these software systems could automatically adapt to hardware changes, it would significantly reduce the maintenance cost and enable rapid upgrade. In this paper, we study the problem of how to automatically adapt to sensor changes, as an important step towards building such long-lived, survivable software systems. We address the adaptation scenarios where a set of sensors are replaced by new sensors. Our approach reconstructs sensor values of replaced sensors by preserving distributions of sensor values before and after the sensor change, thereby not warranting a change in higher-layer software. Compared to existing work, our approach has the following advantages: a) exploiting new sensors without requiring an overlapping period of time between new sensors and old ones; b) providing an estimation of adaptation quality; c) scaling to a large number of sensors. Experiments on weather data and Unmanned Undersea Vehicle (UUV) data demonstrate that our approach can automatically adapt to sensor changes with higher accuracy compared to baseline methods.

Recent progress in AI and Reinforcement learning has shown great success in solving complex problems with high dimensional state spaces. However, most of these successes have been primarily in simulated environments where failure is of little or no consequence. Most real-world applications, however, require training solutions that are safe to operate as catastrophic failures are inadmissible especially when there is human interaction involved. Currently, Safe RL systems use human oversight during training and exploration in order to make sure the RL agent does not go into a catastrophic state. These methods require a large amount of human labor and it is very difficult to scale up. We present a hybrid method for reducing the human intervention time by combining model-based approaches and training a supervised learner to to improve sample efficiency while also ensuring safety. We evaluate these methods on various grid-world environments using both standard and visual representations and show that our approach achieves better performance in terms of sample efficiency, number of catastrophic states reached as well as overall task performance compared to traditional model-free approaches.

Recent successes in deep learning have led to explorations of what makes these techniques so powerful. One goal of such studies is to determine whether such properties can be transferred to alternative learning methods and yield similar benefits. Since the generalization power of any learning algorithm depends upon the inductive bias(es) of that algorithm, we hypothesize that utilizing a bias incorporated by CNNs and other deep methods--spatial locality--can benefit other learning methods as well. We test this hypothesis by incorporating spatial structure when constructing random forests. Our experiments demonstrate that incorporating a spatial locality bias improves the performance of random forests on several image classification tasks.