It's worth noting that dataset this comes from is actually pretty reasonable: "The speech data used are a subset of the MULTEXT multilingual corpus (Campione &

It's worth noting that dataset this comes from is actually pretty reasonable: "The speech data used are a subset of the MULTEXT multilingual corpus (Campione & Véronis 1998, Komatsu et al. 2004). This subset consists of K = 20 texts composed in British English, freely translated into the following languages to convey a comparable semantic content: French (FR), German (GE), Italian (IT), Japanese (JA), Mandarin Chinese (MA), and Spanish (SP). Each text consists of five semantically connected sentences that compose either a narration or a query (to order food by phone, for example)." Which is to say that the study took a set data (constant information content), translated into various languages, and then analysed the information density based on how many syllabels where in it and how long it took to deliver the information (read) by various native speakers. The result being, basically, languages tend toward a common rate of information through-put using a trade off between speed of delivery and data complexity (ie. fewer syllables to communicate an idea -> the language tends to be spoker more slowly). This is pretty interesting stuff; love to see some larger follow up studies. To be honest, I'm not surprised at all to hear japanese has a lower information density, especially in formal stuff; I bet that if you tracked segments in the languages that had _no information content_ you'd find japanese right up there. That said, the blog post doesn't really do much for the article, and just ends up sounding unsubstanciated.