We consider the problem of generating natural language given a communicative goal and a world description.We ask the question: is it possible to combine complementary meaning representations to scale a goal-directed NLG system without losing expressiveness?In particular, we consider using two meaning representations, one based on logical semantics and the other based on distributional semantics.We build upon an existing goal-directed generation system, S-STRUCT, which models sentence generation as planning in a Markov decision process.We develop a hybrid approach, which uses distributional semantics to quickly and imprecisely add the main elements of the sentence and then uses first-order logic based semantics to more slowly add the precise details.We find that our hybrid method allows S-STRUCT's generation to scale significantly better in early phases of generation and that the hybrid can often generate sentences with the same quality as S-STRUCT in substantially less time.However, we also observe and give insight into cases where the imprecision in distributional semantics leads to generation that is not as good as using pure logical semantics.

The use of crowdworkers in NLP research is growing rapidly, in tandem with the exponential increase in research production in machine learning and AI.Ethical discussion regarding the use of crowdworkers within the NLP research community is typically confined in scope to issues related to labor conditions such as fair pay.We draw attention to the lack of ethical considerations related to the various tasks performed by workers, including labeling, evaluation, and production.We find that the Final Rule, the common ethical framework used by researchers, did not anticipate the use of online crowdsourcing platforms for data collection, resulting in gaps between the spirit and practice of human-subjects ethics in NLP research.We enumerate common scenarios where crowdworkers performing NLP tasks are at risk of harm.We thus recommend that researchers evaluate these risks by considering the three ethical principles set up by the Belmont Report.We also clarify some common misconceptions regarding the Institutional Review Board (IRB) application.We hope this paper will serve to reopen the discussion within our community regarding the ethical use of crowdworkers.

This study seeks to develop and test a theory-driven model of technology addiction within the context of smartphone use. Drawing on incentive-sensitisation theory, we proposed a nomological network that centres on a second-order factor of smartphone addiction from a psychological perspective. We empirically evaluated the proposed model against longitudinal data collected from 441 smartphone users. Our analysis from structural equation modelling indicates that our two-factor measurement of smartphone addiction (heightened desire and mobile social interaction) is a parsimonious alternative to the four-factor model. In addition, our findings suggest that addiction and its negative consequences differ in their effects on subsequent perceptions, such as perceived usefulness and perceived enjoyment. We also found that personal innovativeness in information technology and a perception of loneliness explain a large amount of variation in smartphone addiction. Overall, the present study significantly contributes to the information systems literature by offering a refined model of smartphone addiction that effectively describes pathological smartphone use while correcting critical biases in existing models.

Sarcasm detection is an important task in affective computing, requiring large amounts of labeled data.We introduce reactive supervision, a novel data collection method that utilizes the dynamics of online conversations to overcome the limitations of existing data collection techniques.We use the new method to create and release a first-of-its-kind large dataset of tweets with sarcasm perspective labels and new contextual features.The dataset is expected to advance sarcasm detection research.Our method can be adapted to other affective computing domains, thus opening up new research opportunities.

Human communication is multimodal and includes elements such as gesture and facial expression along with spoken language. Modern technology makes it feasible to capture all such aspects of communication in natural settings. As a result, similar to fields such as genetics, astronomy and neuroscience, scholars in areas such as linguistics and communication studies are on the verge of a data-driven revolution in their fields. These new approaches require analytical support from machine learning and artificial intelligence to develop tools to help process the vast data repositories. The Distributed Little Red Hen Lab project is an international team of interdisciplinary researchers building a large-scale infrastructure for data-driven multimodal communications research. In this paper, we describe a machine learning system developed to automatically annotate a large database of television program videos as part of this project. The annotations mark regions where people or speakers are on screen along with body part motions including head, hand and shoulder motion. We also annotate a specific class of gestures known as timeline gestures. An existing gesture annotation tool, ELAN, can be used with these annotations to quickly locate gestures of interest. Finally, we provide an update mechanism for the system based on human feedback. We empirically evaluate the accuracy of the system as well as present data from pilot human studies to show its effectiveness at aiding gesture scholars in their work.

The goal of the IEEE 2018 FEMH Voice Data Challenge was to develop an effective algorithmic approach to classifying voice samples as normal or pathological, and further subdivide the pathological samples into three types. We adopted a multi-representation ensemble approach to the task. We designed a pipeline with three classification stages, where each stage used a combination of supervised, semi-supervised and multiple-instance learners. This approach was able to achieve a sensitivity of 89% and specificity of 76% in classifying normal from pathological samples and an unweighted average recall (UAR) of 60.67% in subclassifying pathological samples into three types.

An interesting observation about the well-known AdaBoost algorithm is that, though theory suggests it should overfit when applied to noisy data, experiments indicate it often does not do so in practice. In this paper, we study the behavior of AdaBoost on datasets with one-sided uniform class noise using linear classifiers as the base learner. We show analytically that, under some ideal conditions, this approach will not overfit, and can in fact recover a zero-error concept with respect to the true, uncorrupted instance labels. We also analytically show that AdaBoost increases the margins of predictions over boosting iterations, as has been previously suggested in the literature. We then compare the empirical behavior of AdaBoost using real world datasets with one-sided noise derived from multiple-instance data. Although our assumptions may not hold in a practical setting, our experiments show that standard AdaBoost still performs well, as suggested by our analysis, and often outperforms baseline variations in the literature that explicitly try to account for noise.

An estimated 17.5 million people died from a cardiovascular disease in 2012, representing 31% of all global deaths. Most acute coronary events result from rupture of the protective fibrous cap overlying an atherosclerotic plaque. The task of early identification of plaque types that can potentially rupture is, therefore, of great importance. The state-ofthe-art approach to imaging blood vessels is intravascular optical coherence tomography (IVOCT). However, currently, this is an offline approach where the images are first collected and then manually analyzed a frame at a time to identify regions at risk of thrombosis. This process is extremely laborious, time consuming and prone to human error. We are building a system that, when complete, will provide interactive 3D visualization of a blood vessel as an IVOCT is in progress. The visualization will highlight different plaque types and enable quick identification of regions at risk for thrombosis. In this paper, we describe our approach, focusing on machine learning methods that are a key enabling technology. Our empirical results using real OCT data show that our approach can identify different plaque types efficiently with high accuracy across multiple patients.

A key goal in natural language generation (NLG) is to enable fast generation even with large vocabularies, grammars and worlds.In this work, we build upon a recently proposed NLG system, Sentence Tree Realization with UCT (STRUCT).We describe four enhancements to this system: (i) pruning the grammar based on the world and the communicative goal, (ii) intelligently caching and pruning the combinatorial space of semantic bindings, (iii) reusing the lookahead search tree at different search depths, and (iv) learning and using a search control heuristic.We evaluate the resulting system on three datasets of increasing size and complexity, the largest of which has a vocabulary of about 10K words, a grammar of about 32K lexicalized trees and a world with about 11K entities and 23K relations between them.Our results show that the system has a median generation time of 8.5s and finds the best sentence on average within 25s.These results are based on a sequential, interpreted implementation and are significantly better than the state of the art for planningbased NLG systems.

An important question in feature selection is whether a selection strategy recovers the “true” set of features, given enough data. We study this question in the context of the popular Least Absolute Shrinkage and Selection Operator (Lasso) feature selection strategy. In particular, we consider the scenario when the model is misspecified so that the learned model is linear while the underlying real target is nonlinear. Surprisingly, we prove that under certain conditions, Lasso is still able to recover the correct features in this case. We also carry out numerical studies to empirically verify the theoretical results and explore the necessity of the conditions under which the proof holds.

We propose a new theoretical framework for analyzing the multiple-instance learning (MIL) setting. In MIL, training examples are provided to a learning algorithm in the form of labeled sets, or "bags," of instances. Applications of MIL include 3-D quantitative structure--activity relationship prediction for drug discovery and content-based image retrieval for web search. The goal of an algorithm is to learn a function that correctly labels new bags or a function that correctly labels new instances. We propose that bags should be treated as latent distributions from which samples are observed. We show that it is possible to learn accurate instance- and bag-labeling functions in this setting as well as functions that correctly rank bags or instances under weak assumptions. Additionally, our theoretical results suggest that it is possible to learn to rank efficiently using traditional, well-studied "supervised" learning approaches. We perform an extensive empirical evaluation that supports the theoretical predictions entailed by the new framework. The proposed theoretical framework leads to a better understanding of the relationship between the MI and standard supervised learning settings, and it provides new methods for learning from MI data that are more accurate, more efficient, and have better understood theoretical properties than existing MI-specific algorithms.

We analyze and evaluate a generative process for multiple-instance learning (MIL) in which bags are distributions over instances. We show that our generative process contains as special cases generative models explored in prior work, while excluding scenarios known to be hard for MIL. Further, under the mild assumption that every negative instance is observed with nonzero probability in some negative bag, we show that it is possible to learn concepts that accurately label instances from MI data in this setting. Finally, we show that standard supervised approaches can learn concepts with low area-under-ROC error from MI data in this setting. We validate this surprising result with experiments using several synthetic and real-world MI datasets that have been annotated with instance labels.

We study the problem of generating an English sentence given an underlying probabilistic grammar, a world and a communicative goal.We model the generation problem as a Markov decision process with a suitably defined reward function that reflects the communicative goal.We then use probabilistic planning to solve the MDP and generate a sentence that, with high probability, accomplishes the communicative goal.We show empirically that our approach can generate complex sentences with a speed that generally matches or surpasses the state of the art.Further, we show that our approach is anytime and can handle complex communicative goals, including negated goals.

Resampling techniques such as bagging are often used in supervised learning to produce more accurate classifiers. In this work, we show that multiple-instance learning admits a different form of resampling, which we call "shuffling." In shuffling, we resample instances in such a way that the resulting bags are likely to be correctly labeled. We show that resampling results in both a reduction of bag label noise and a propagation of additional informative constraints to a multiple-instance classifier. We empirically evaluate shuffling in the context of multiple-instance classification and multiple-instance active learning and show that the approach leads to significant improvements in accuracy.

Adverse and anomalous (A&A) events are a serious concern in medical robots. We describe a system that can rapidly detect such events and predict their occurrence. As part of this system, we describe simulation, data collection and user interface tools we build for a robot for small animal biopsies. The data we collect consists of both the hardware state of the robot and variables in the software controller. We use this data to train dynamic Bayesian network models of the joint hardware-software state-space dynamics of the robot. Our empirical evaluation shows that (i) our models can accurately model normal behavior of the robot, (ii) they can rapidly detect anomalous behavior once it starts, (iii) they can accurately predict a future A&A event within a time window of it starting and (iv) the use of additional software variables beyond the hardware state of the robot is important in being able to detect and predict certain kinds of events.

In the context of field testing (operational testing) of software, we address the problem of balancing the potential reduction in failure risk that developers may achieve by reviewing captured test executions to identify failures (and by successfully debugging their causes) against the cost of reviewing the tests. To achieve a desirable balance, we propose a cost-sensitive active learning strategy. Our approach guides developers in selecting a sample of test executions to review and label, and it calls for them to profile execution dynamics and characterize the symptoms and relative severity levels of failures. Profiles, labels, failure symptoms, and severity levels are used by the active learner to construct and refine a mapping between examined and unexamined tests, on one hand, and possible defects, on the other hand. This mapping is used together with estimates of test review costs to guide the selection of additional tests. We evaluate our approach on three subject programs and show that it (1) produces reasonable predictions of risk reduction and (2) significantly improves severity-weighted reliability for each subject program, with relatively low developer effort.

We describe an approach to incorporating Bayesian priors in the maxq framework for hierarchical reinforcement learning (HRL). We define priors on the primitive environment model and on task pseudo-rewards. Since models for composite tasks can be complex, we use a mixed model-based/model-free learning approach to find an optimal hierarchical policy. We show empirically that (i) our approach results in improved convergence over non-Bayesian baselines, given sensible priors, (ii) task hierarchies and Bayesian priors can be complementary sources of information, and using both sources is better than either alone, (iii) taking advantage of the structural decomposition induced by the task hierarchy significantly reduces the computational cost of Bayesian reinforcement learning and (iv) in this framework, task pseudo-rewards can be learned instead of being manually specified, leading to automatic learning of hierarchically optimal rather than recursively optimal policies.

MOTIVATION: Several software tools specialize in the alignment of short next-generation sequencing reads to a reference sequence. Some of these tools report a mapping quality score for each alignment-in principle, this quality score tells researchers the likelihood that the alignment is correct. However, the reported mapping quality often correlates weakly with actual accuracy and the qualities of many mappings are underestimated, encouraging the researchers to discard correct mappings. Further, these low-quality mappings tend to correlate with variations in the genome (both single nucleotide and structural), and such mappings are important in accurately identifying genomic variants. APPROACH: We develop a machine learning tool, LoQuM (LOgistic regression tool for calibrating the Quality of short read mappings, to assign reliable mapping quality scores to mappings of Illumina reads returned by any alignment tool. LoQuM uses statistics on the read (base quality scores reported by the sequencer) and the alignment (number of matches, mismatches and deletions, mapping quality score returned by the alignment tool, if available, and number of mappings) as features for classification and uses simulated reads to learn a logistic regression model that relates these features to actual mapping quality. RESULTS: We test the predictions of LoQuM on an independent dataset generated by the ART short read simulation software and observe that LoQuM can 'resurrect' many mappings that are assigned zero quality scores by the alignment tools and are therefore likely to be discarded by researchers. We also observe that the recalibration of mapping quality scores greatly enhances the precision of called single nucleotide polymorphisms. AVAILABILITY: LoQuM is available as open source at http://compbio.case.edu/loqum/. CONTACT: [email protected].

Previous work has shown that application of graph mining techniques to system dependence graphs improves the precision of automatic defect discovery by revealing subgraphs corresponding to implicit programming rules and to rule violations. However, developers must still confirm, edit, or discard reported rules and violations, which is both costly and error-prone. In order to reduce developer effort and further improve precision, we investigate the use of supervised learning models for classifying and ranking rule and violation subgraphs. In particular, we present and evaluate logistic regression models for rules and violations, respectively, which are based on general dependence-graph features. Our empirical results indicate that (i) use of these models can significantly improve the precision and recall of defect discovery, and (ii) our approach is superior to existing heuristic approaches to rule and violation ranking and to an existing static-warning classifier, and (iii) accurate models can be learned using only a few labeled examples.

We present an algorithm, HI-MAT (Hierarchy Induction via Models And Trajectories), that discovers MAXQ task hierarchies by applying dynamic Bayesian network models to a successful trajectory from a source reinforcement learning task. HI-MAT discovers subtasks by analyzing the causal and temporal relationships among the actions in the trajectory. Under appropriate assumptions, HI-MAT induces hierarchies that are consistent with the observed trajectory and have compact value-function tables employing safe state abstractions. We demonstrate empirically that HI-MAT constructs compact hierarchies that are comparable to manually-engineered hierarchies and facilitate significant speedup in learning when transferred to a target task.

We present a new machine learning approach for 3D-QSAR, the task of predicting binding affinities of molecules to target proteins based on 3D structure. Our approach predicts binding affinity by using regression on substructures discovered by relational learning. We make two contributions to the state-of-the-art. First, we use multiple-instance (MI) regression, which represents a molecule as a set of 3D conformations, to model activity. Second, the relational learning component employs the "Score As You Use" (SAYU) method to select substructures for their ability to improve the regression model. This is the first application of SAYU to multiple-instance, real-valued prediction. We evaluate our approach on three tasks and demonstrate that (i) SAYU outperforms standard coverage measures when selecting features for regression, (ii) the MI representation improves accuracy over standard single feature-vector encodings and (iii) combining SAYU with MI regression is more accurate for 3D-QSAR than either approach by itself.

We consider the problem of multi-task reinforcement learning, where the agent needs to solve a sequence of Markov Decision Processes (MDPs) chosen randomly from a fixed but unknown distribution. We model the distribution over MDPs using a hierarchical Bayesian infinite mixture model. For each novel MDP, we use the previously learned distribution as an informed prior for modelbased Bayesian reinforcement learning. The hierarchical Bayesian framework provides a strong prior that allows us to rapidly infer the characteristics of new environments based on previous environments, while the use of a nonparametric model allows us to quickly adapt to environments we have not encountered before. In addition, the use of infinite mixtures allows for the model to automatically learn the number of underlying MDP components. We evaluate our approach and show that it leads to significant speedups in convergence to an optimal policy after observing only a small number of tasks.

In a multiple instance (MI) learning problem, instances are naturally organized into bags and it is the bags, instead of individual instances, that are labeled for training. MI learners assume that every instance in a bag labeled negative is actually negative, whereas at least one instance in a bag labeled positive is actually positive. We present a framework for active learning in the multiple-instance setting. In particular, we consider the case in which an MI learner is allowed to selectively query unlabeled instances in positive bags. This approach is well motivated in domains in which it is inexpensive to acquire bag labels and possible, but expensive, to acquire instance labels. We describe a method for learning from labels at mixed levels of granularity, and introduce two active query selection strategies motivated by the MI setting. Our experiments show that learning from instance labels can significantly improve performance of a basic MI learning algorithm in two multiple-instance domains: content-based image recognition and text classification.

We empirically study the relationship between supervised and multiple instance (MI) learning. Algorithms to learn various concepts have been adapted to the MI representation. However, it is also known that concepts that are PAC-learnable with one-sided noise can be learned from MI data. A relevant question then is: how well do supervised learners do on MI data? We attempt to answer this question by looking at a cross section of MI data sets from various domains coupled with a number of learning algorithms including Diverse Density, Logistic Regression, nonlinear Support Vector Machines and FOIL. We consider a supervised and MI version of each learner. Several interesting conclusions emerge from our work: (1) no MI algorithm is superior across all tested domains, (2) some MI algorithms are consistently superior to their supervised counterparts, (3) using high false-positive costs can improve a supervised learner's performance in MI domains, and (4) in several domains, a supervised algorithm is superior to any MI algorithm we tested.

This paper extends previous work on skewing, an approach to problematic functions in decision tree induction. The previous algorithms were applicable only to functions of binary variables. In this paper, we extend skewing to directly handle functions of continuous and nominal variables. We present experiments with randomly generated functions and a number of real world datasets to evaluate the algorithm's accuracy. Our results indicate that our algorithm almost always outperforms an Information Gain-based decision tree learner.

We analyze skewing, an approach that has been empirically observed to enable greedy decision tree learners to learn "difficult" Boolean functions, such as parity, in the presence of irrelevant variables. We prove tha, in an idealized setting, for any function and choice of skew parameters, skewing finds relevant variables with probability 1. We present experiments exploring how different parameter choices affect the success of skewing in empirical settings. Finally, we analyze a variant of skewing called Sequential Skewing.

BACKGROUND: The BioCreative text mining evaluation investigated the application of text mining methods to the task of automatically extracting information from text in biomedical research articles. We participated in Task 2 of the evaluation. For this task, we built a system to automatically annotate a given protein with codes from the Gene Ontology (GO) using the text of an article from the biomedical literature as evidence. METHODS: Our system relies on simple statistical analyses of the full text article provided. We learn n-gram models for each GO code using statistical methods and use these models to hypothesize annotations. We also learn a set of Naïve Bayes models that identify textual clues of possible connections between the given protein and a hypothesized annotation. These models are used to filter and rank the predictions of the n-gram models. RESULTS: We report experiments evaluating the utility of various components of our system on a set of data held out during development, and experiments evaluating the utility of external data sources that we used to learn our models. Finally, we report our evaluation results from the BioCreative organizers. CONCLUSION: We observe that, on the test data, our system performs quite well relative to the other systems submitted to the evaluation. From other experiments on the held-out data, we observe that (i) the Naïve Bayes models were effective in filtering and ranking the initially hypothesized annotations, and (ii) our learned models were significantly more accurate when external data sources were used during learning.

This paper extends previous work on the Skewing algorithm, a promising approach that allows greedy decision tree induction algorithms to handle problematic functions such as parity functions with a lower run-time penalty than Lookahead. A deficiency of the previously proposed algorithm is its inability to scale up to high dimensional problems. In this paper, we describe a modified algorithm that scales better with increasing numbers of variables. We present experiments with randomly generated Boolean functions that evaluate the algorithm's response to increasing dimensions. We also evaluate the algorithm on a challenging real world biomedical problem, that of SH3 domain binding. Our results indicate that our algorithm almost always outperforms an information gain-based decision tree learner.