To estimate the degree of sincerity conveyed by a speech utterance and received by listeners, we propose an instance-based learning framework with shallow neural networks. The framework plays as not only a regressor that intends to fit the predicted value to the actual value but also a ranker that preserves the relative target magnitude between each pair of utterances, in an attempt to derive a higher Spearman's rank correlation coefficient. In addition to describing how to simultaneously minimize regression and ranking losses, the issue of how utterance pairs work in the training and evaluation phases is also addressed by two kinds of realizations. The intuitive one is related to random sampling while the other seeks for representative utterances, named anchors, to form non-stochastic pairs. Our system outperforms the baseline by more than 25% relative improvement in the development set.
The development of an automatic oral presentation assessment system is important for the educational researchers to assess and train the communication ability of school leaders. In this work, we aim at enhancing the performance of the existing pre-service school principals' presentation scoring system by including lexical information as an additional modality. We propose to use latent n-grams distributed word representations and weighted counts of part-of-speech tag to derive features from the speech transcripts in the National Academy for Educational Research (NAER) oral presentation database. We carry out two different experiments: Exp I is a binary classification task between high versus low performing speech, and Exp II is a continuous scoring on the entire dataset. In Exp I, the proposed framework achieves a competitive accuracy of 0.79, and in Exp II, by fusing this text-based system to the existing audio-video based system, we obtain a spearman correlation of 0.641 (18.05% relative improvement). The two experiments demonstrate the modeling power of our proposed framework and signify the substantial complementary information in the lexical modality while assessing the quality of an oral presentation.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.