Latent variable-based multiple instance learning towards label-free polarity detection
Peter Lajos Ihasz, Mate Kovacs, Victor V. Kryssanov · 2019
Extracting information from the audio content of the users' dialogic utterances would provide an easily-perturbed set of features that could serve as a reliable and inexpensive mean for emotion recognition, suitable to be applied in commercial software development. Owing to the diversity of audio features, however, emotion recognition in spontaneous dialogues is a complex task, typically requiring the pre-training of classifiers on large collections of labeled data. To escape the necessity of hand labeling, a novel multiple instance learning method is proposed. It performs the bag-label-based instance classification through the extraction of latent variables with variational autoencoders. In this semi-supervised method, the bags themselves are gathered from audio features of "weakly" labeled YouTube videos, thus training is fully automated and does not require manual annotation.