State-of-the-art named entity recognition systems rely heavily on hand-crafted features and domain-specific knowledge in order to learn effectively from the small, supervised training corpora that are available. In this paper, we introduce two new neural architectures-one based on bidirectional LSTMs and conditional random fields, and the other that constructs and labels segments using a transition-based approach inspired by shift-reduce parsers. Our models rely on two sources of information about words: character-based word representations learned from the supervised corpus and unsupervised word representations learned from unannotated corpora. Our models obtain state-of-the-art performance in NER in four languages without resorting to any language-specific knowledge or resources such as gazetteers. 1

No abstract

Background and Purpose-Upper limb motor impairment poststroke is commonly evaluated using clinical outcome measures such as the Fugl-Meyer Assessment. However, most clinical measures provide little information about motor patterns and compensations (eg, trunk displacement) used for task performance. Such information is obtained using movement quality kinematic variables (joint ranges, trunk displacement). Evaluation of movement quality may also help distinguish between levels of motor impairment severity in individuals poststroke. Our objective was to estimate concurrent and discriminant validity of movement quality kinematic variables for pointing and reach-to-grasp tasks. Methods-A retrospective study of kinematic data (sagittal trunk displacement, shoulder flexion, shoulder horizontal adduction, elbow extension) and Fugl-Meyer Assessment scores from 86 subjects (subacute to chronic stroke) performing pointing and reaching tasks was done. Multiple and logistic regression analyses were used to estimate concurrent and discriminant validity respectively. Cutoff points for distinguishing between levels of upper limb motor impairment severity (mild, moderate to severe) were estimated using sensitivity/specificity decision plots. The criterion measure used was the Fugl-Meyer Assessment (upper limb section). Results-The majority of variance in Fugl-Meyer Assessment scores was explained by a combination of trunk displacement and shoulder flexion (51%) for the pointing task and by trunk displacement alone (52%) for the reach-to-grasp task. Trunk displacement was the only variable that distinguished between levels of motor impairment severity. Cutoff points were 4.8 cm for pointing and 10.2 cm for reach-to-grasp movements. Conclusion-Movement

The results suggest that people with stroke may be capable of using extrinsic feedback for implicit motor learning and improving UL motor recovery. Emergent questions regarding the advantages of using different media for feedback delivery and the optimal type and schedule of feedback to enhance motor learning in patient populations still need to be addressed.

Introduction. Despite interest in virtual environments (VEs) for poststroke arm motor rehabilitation, advantages over physical environment (PE) training have not been established. Objective. The authors compared kinematic and clinical outcomes of dose-matched upper-limb training between a 3D VE and a PE in chronic stroke. Methods: Participants (n = 32) were randomized to a 3D VE or PE for training. They pointed to 6 workspace targets (72 trials, 12 trials/target, randomized) for 12 sessions over 4 weeks with similar feedback on precision, movement speed, and trunk displacement. Primary (kinematics, clinical arm motor impairment) and secondary (activity level, arm use) outcomes were compared by time (PRE, POST, and follow-up, RET), training environment, and impairment severity (mild, moderate-to-severe) using mixed-model analyses of variance (ANOVAs). Results. Endpoint speed, overall performance on a reach-to-grasp task, and activity levels increased in both groups. Only participants in the VE group improved shoulder horizontal adduction at POST (9.5°) and flexion at both POST (6.3°) and RET (13°). Impairment level affected outcomes. After VE training, the mild group increased elbow extension (RET, 25.5°). The moderate-to-severe group in VE increased arm use at POST (0.5 points) and reaching ability at RET (2.2 points). The moderate-to-severe group training in PE increased reaching ability earlier (POST, 1.7 points) and both elbow extension (10.7°) and arm use (0.4 points) at RET, but these changes were accompanied by increased compensatory trunk displacement (RET, 30.2 mm). Conclusion. VE training led to more changes in the mild group and a motor recovery pattern in the moderate-to-severe group indicative of less compensation, possibly because of a better use of feedback.

We propose a recurrent neural model that generates natural-language questions from documents, conditioned on answers. We show how to train the model using a combination of supervised and reinforcement learning. After teacher forcing for standard maximum likelihood training, we fine-tune the model using policy gradient techniques to maximize several rewards that measure question quality. Most notably, one of these rewards is the performance of a questionanswering system. We motivate question generation as a means to improve the performance of question answering systems. Our model is trained and evaluated on the recent question-answering dataset SQuAD.

Neural generative models have been become increasingly popular when building conversational agents. They offer flexibility, can be easily adapted to new domains, and require minimal domain engineering. A common criticism of these systems is that they seldom understand or use the available dialog history effectively. In this paper, we take an empirical approach to understanding how these models use the available dialog history by studying the sensitivity of the models to artificially introduced unnatural changes or perturbations to their context at test time. We experiment with 10 different types of perturbations on 4 multi-turn dialog datasets and find that commonly used neural dialog architectures like recurrent and transformer-based seq2seq models are rarely sensitive to most perturbations such as missing or reordering utterances, shuffling words, etc. Also, by open-sourcing our code, we believe that it will serve as a useful diagnostic tool for evaluating dialog systems in the future 1 .

Purpose: Motor and cognitive impairments are common and often coexist in patients with stroke. Although evidence is emerging about specific relationships between cognitive deficits and upper-limb motor recovery, the practical implication of these relationships for rehabilitation is unclear. Using a structured review and meta-analyses, we examined the nature and strength of the associations between cognitive deficits and upper-limb motor recovery in studies of patients with stroke.Methods: Motor recovery was defined using measures of upper limb motor impairment and/or activity limitations. Studies were included if they reported on at least one measure of cognitive function and one measure of upper limb motor impairment or function.Results: Six studies met the selection criteria. There was a moderate association (r = 0.43; confidence interval; CI:0.09– 0.68, p = 0.014) between cognition and overall arm motor recovery. Separate meta-analyses showed a moderately strong association between executive function and motor recovery (r = 0.48; CI:0.26– 0.65; p <  0.001), a weak positive correlation between attention and motor recovery (r = 0.25; CI:0.04– 0.45; p = 0.023), and no correlation between memory and motor recovery (r = 0.42; CI:0.16– 0.79; p = 0.14).Conclusion: These results imply that information on the presence of cognitive deficits should be considered while planning interventions for clients in order to design more personalized interventions tailored to the individual for maximizing upper-limb recovery.