Alarge-scaleparallelcorpusisindispensabletotrainencoder-decoderneuralmachinetranslation. Themethod of using synthetic parallel texts, called back-translation, in which target monolingual sentences are automatically translated into the source language, has been proven effective in improving the decoder. However, it does not necessarily help enhance the encoder. In this paper, we propose a method that enhances not only the decoder but also the encoder using target monolingual corpora by generating multiple source sentences via sampling-based sequence generation. The source sentences generated in that way increase their diversity and thus help make the encoder robust. Ourexperimentalresultsshowthatthetranslationqualitywasimprovedbyincreasingthenumberofsynthetic source sentences for each given target sentence. Even though the quality did not reach to the one that realized with a genuine parallel corpus comprising single human translations, our proposed method derived over 50% of the improvementbroughtbytheparallelcorpususingonlyitstargetside, i.e., monolingualdata. Moreover,theproposed samplingmethodresultedinﬁnaltranslationofhigherqualitythann-bestback-translation. Theseresultsindicatethat not only the quality of back-translation but also the diversity of synthetic source sentences is crucial.
In recent years, population with autism spectrum disorder (ASD) are growing explosively, and diagnosis of ASD is difficult due to difference of interviewers and environments, etc. We show relations between utterance features and ASD severity scores, which were manually given by a clinical psychologist. These scores are of the Autism Diagnostic Observation Schedule (ADOS), which is one of the standard metrics for symptom evaluation for subjects who are suspected as ASD. We built our original corpus where we transcribed voice records of our ADOS evaluation experiment movies. Our corpus is the world largest as speech/dialog of ASD subjects, and there has been no such ADOS corpus available in Japanese language as far as we know. We investigated relationships between ADOS scores (severity) and utterance features we defined. Our system automatically estimated their scores using support vector regression (SVR). Our average estimation errors were around error rates that human ADOS experts are required not to exceed. Because our detailed analysis for each part of the ADOS test (“puzzle toy assembly + story telling” part and the “depiction of a picture” part) shows different error rates, effectiveness of our features would depend on the contents of the records. By comparing an ADOS score prediction result of adults and adults with that of children, we showed common features of ADOS scores between children and adults. Our entire results suggest a new automatic way to assist humans’ diagnosis, which could help supporting language rehabilitation for patients with ASD in future.
Residual Networks with convolutional layers are widely used in the field of machine learning. Since they effectively extract features from input data by stacking multiple layers, they can achieve high accuracy in many applications. However, the stacking of many layers raises their computation costs. To address this problem, we propose Network Implosion, it erases multiple layers from Residual Networks without degrading accuracy. Our key idea is to introduce a priority term that identifies the importance of a layer; we can select unimportant layers according to the priority and erase them after the training. In addition, we retrain the networks to avoid critical drops in accuracy after layer erasure. Our experiments show that Network Implosion can, for classification on CIFAR10/100 and ImageNet, reduce the number of layers by 24.00% ～ 42.86% without any drop in accuracy.
This paper implements the infection process of 2019 Novel Coronavirus Diseases (COVID-19) in an agentbasedmodel and compares the effectiveness of multiple infection prevention measures. In the model, 1120 virtualresidents agents live in two towns where they commute to office or school and visiting stores. The model simulates aninfection process in which they were exposed to the risk of transmission of the novel coronavirus. The results of theexperiments showed that individual infection prevention measures (commuting, teleworking, class closing, contactrate reduction, staying at home after fever) alone or partially combined them do not produce significant effects. Onthe other hand, if comprehensive measures were taken, it was confirmed that the number of deaths, the infectionrate, and the number of severe hospitalised patients per day were decreased significantly at the median and maximumrespectively.