AmericasNLI: Machine translation and natural language inference systems for Indigenous languages of the Americas.

Little attention has been paid to the development of human language technology for truly low-resource languages-i.e., languages with limited amounts of digitally available text data, such as Indigenous languages. However, it has been shown that pretrained multilingual models are able to perform crosslingual transfer in a zero-shot setting even for low-resource languages which are unseen during pretraining. Yet, prior work evaluating performance on unseen languages has largely been limited to shallow token-level tasks. It remains unclear if zero-shot learning of deeper semantic tasks is possible for unseen languages. To explore this question, we present AmericasNLI, a natural language inference dataset covering 10 Indigenous languages of the Americas. We conduct experiments with pretrained models, exploring zero-shot learning in combination with model adaptation. Furthermore, as AmericasNLI is a multiway parallel dataset, we use it to benchmark the performance of different machine translation models for those languages. Finally, using a standard transformer model, we explore translation-based approaches for natural language inference. We find that the zero-shot performance of pretrained models without adaptation is poor for all languages in AmericasNLI, but model adaptation via continued pretraining results in improvements. All machine translation models are rather weak, but, surprisingly, translation-based approaches to natural language inference outperform all other models on that task.

Automated social skills training with audiovisual information.

People with social communication difficulties tend to have superior skills using computers, and as a result computer-based social skills training systems are flourishing. Social skills training, performed by human trainers, is a well-established method to obtain appropriate skills in social interaction. Previous works have attempted to automate one or several parts of social skills training through human-computer interaction. However, while previous work on simulating social skills training considered only acoustic and linguistic features, human social skills trainers take into account visual features (e.g. facial expression, posture). In this paper, we create and evaluate a social skills training system that closes this gap by considering audiovisual features regarding ratio of smiling, yaw, and pitch. An experimental evaluation measures the difference in effectiveness of social skill training when using audio features and audiovisual features. Results showed that the visual features were effective to improve users' social skills.

Removing noise from event-related potentials using a probabilistic generative model with grouped covariance matrices.

Analysis of electroencephalograms (EEG) usually suffers from a variety of noises. In this paper, we propose a new method for background noise removal from single-trial event-related potentials (ERPs) recorded with a multi-channel EEG. An observed signal is separated into multiple signals with a multi-channel Wiener filter, whose coefficients are estimated based on a probabilistic generative model in the time-frequency domain. The main contribution is a method to estimate covariance matrices for each frequency bins of short-time Fourier transform (STFT) representing different spatial spread of a multi-channel EEG signal according to frequencies. An experiment using a pseudo-ERP data set demonstrates the effectiveness of our proposed method.

An evaluation of EEG ocular artifact removal with a multi-channel wiener filter based on probabilistic generative model.

Data contamination by ocular artifacts such as eye blinks and eye movements is a major barrier that must be overcome when attempting to analyze electroencephalogram (EEG) and event-related potential (ERP) data. To handle this problem, a number of artifact removal methods has been proposed. Specifically, we focus on a method using a multi-channel Wiener filters based on a probabilistic generative model. This method assumes that the observed signal is the sum of multiple signals elicited by psychological or physical events, and separates the observed signal into each event signal using estimated model parameters. Based on this scheme, we have proposed a model parameter estimation method using prior information of each event signal. In this paper, we examine the potential of this model to deal with highly contaminated signals by collecting EEG data intentionally contaminated by eye blinks and relatively clean ERP data, and using them as prior information of each event signal. We conducted an experimental evaluation using a classical attention task. The results showed the proposed method effectively enhances the target ERP component while reducing the contamination caused by eye blinks.