Скачать или смотреть Tutorial 12: Word Embeddings for Descriptive Corpus Analysis: Analogies, Polysemy, and Stability

Description:

Word embeddings such as word2vec have recently garnered attention as potentially useful tools for analysis in social science. They promise an unsupervised method to quantify the connotations of words, and compare these across time or different subgroups. However, when training or using word embeddings, researchers may find that they don’t work as well as expected, or produce unreplicable results. We focus on three subtle issues in their use that could result in misleading observations: (1) indiscriminate use of analogical reasoning, which has been shown to underperform on many types of analogies; (2) the surprising prevalence of polysemous words and distributional similarity of antonyms, both leading to counterintuitive results; and (3) instability in nearest-neighbor distances caused by sensitivity to noise in the training process. Through demonstrations, we will learn how to detect, understand, and most importantly mitigate the effects of these issues.

Tutorial Host: Neha Kennard

Slides: bit.ly/embedding-tutorial-slides

Code: bit.ly/embedding-tutorial-notebook

This is part of a larger tutorial series, NLP+CSS 201: Beyond the basics, which is organized by Ian Stewart and Katherine Keith. Website: https://nlp-css-201-tutorials.github....