Automated content analysis across six languages

摘要

Corpus selection bias in international relations research presents an epistemological problem: How do we know what we know? Most social science research in the field of text analytics relies on English language corpora, biasing our ability to understand international phenomena. To address the issue of corpus selection bias, we introduce results that suggest that machine translation may be used to address non-English sources. We use human translation and machine translation (Google Translate) on a collection of aligned sentences from United Nations documents extracted from the Multi-UN corpus, analyzed with a “bag of words” analysis tool, Linguistic Inquiry Word Count (LIWC). Overall, the LIWC indices proved relatively stable across machine and human translated sentences. We find that while there are statistically significant differences between the original and translated documents, the effect sizes are relatively small, especially when looking at psychological processes.