No abstract

For medical students, virtual patient dialogue systems can provide useful training opportunities without the cost of employing actors to portray standardized patients. This work utilizes word-and character-based convolutional neural networks (CNNs) for question identification in a virtual patient dialogue system, outperforming a strong word-and characterbased logistic regression baseline. While the CNNs perform well given sufficient training data, the best system performance is ultimately achieved by combining CNNs with a hand-crafted pattern matching system that is robust to label sparsity, providing a 10% boost in system accuracy and an error reduction of 47% as compared to the pattern-matching system alone.

This paper presents evidence of a linguistic focus effect on coreference resolution in broad-coverage human sentence processing. While previous work has explored the role of prominence in coreference resolution (Almor, 1999; Foraker and McElree, 2007), these studies use constructed stimuli with specific syntactic patterns (e.g. cleft constructions) which could have idiosyncratic frequency confounds. This paper explores the generalizability of this effect on coreference resolution in a broad-coverage analysis. In particular, the current work proposes several new estimators of prominence appropriate for broadcoverage sentence processing and evaluates them as predictors of reading behavior in the Natural Stories corpus (Futrell, Gibson, Tily, Vishnevetsky, Piantadosi, and Fedorenko, in prep), a collection of "constructed-natural" narratives read by a large number of subjects. Results show a strong facilitation effect for one of these predictors on exploratory data and confirm that it generalizes to held-out data. These results provide broad-coverage support for the hypothesis that coreference resolution is easier when the target entity is focused by discourse properties, resulting in faster reading times.

We present a log-linear ranking model for interpreting questions in a virtual patient dialogue system and demonstrate that it substantially outperforms a more typical multiclass classifier model using the same information. The full model makes use of weighted and concept-based matching features that together yield a 15% error reduction over a strong lexical overlap baseline. The accuracy of the ranking model approaches that of an extensively handcrafted pattern matching system, promising to reduce the authoring burden and make it possible to use confidence estimation in choosing dialogue acts; at the same time, the effectiveness of the concept-based features indicates that manual development resources can be productively employed with the approach in developing concept hierarchies.

We survey research using neural sequence-to-sequence models as computational models of morphological learning and learnability. We discuss their use in determining the predictability of inflectional exponents, in making predictions about language acquisition and in modeling language change. Finally, we make some proposals for future work in these areas. 1 introduction Theoretical morphologists have long appealed to notions of learning, or learnability, to explain language change and the varied typological patterns of the world's languages. The high-level argument is simple: all natural languages must be learned, and "unlearnable" linguistic systems cannot survive. Therefore, the learning mechanism provides constraints on what sorts of languages can exist in the world. In the realm of morphology, however, it has not proven simple to define