2018/10/25 by Sven Buechel, João Sedoc, Buechel, Sven +5
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Sentiment Analysis and Opinion Mining #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1810.10949
openalex publication_date 2018/10/25 · openalex created_date 2022/08/02 · openalex updated_date 2026/07/28
One of the major downsides of Deep Learning is its supposed need for vast\namounts of training data. As such, these techniques appear ill-suited for NLP\nareas where annotated data is limited, such as less-resourced languages or\nemotion analysis, with its many nuanced and hard-to-acquire annotation formats.\nWe conduct a questionnaire study indicating that indeed the vast majority of\nresearchers in emotion analysis deems neural models inferior to traditional\nmachine learning when training data is limited. In stark contrast to those\nsurvey results, we provide empirical evidence for English, Polish, and\nPortuguese that commonly used neural architectures can be trained on\nsurprisingly few observations, outperforming n-gram based ridge regression on\nonly 100 data points. Our analysis suggests that high-quality, pre-trained word\nembeddings are a main factor for achieving those results.\n