But it may very well be correctly understanding that people like to end sentences with an emoji, combined with that being the most likely emoji used in conversations like this. That becomes plausible when combined with most such conversations using little to no emojis (more professional and dry) but those that do (more emotive) tend to be between partners. Texting a partner about these topics might be more common than texting others, or at least more commonly emoji-heavy, which would be below the noise floor until also mixing in the probabilities of sentences being followed by emoji versus being followed by any particular word. These particular conversations would tend to express
passionate love more than happiness, sadness, sarcasm, taking orders, or various other emotions. Suddenly a signal emerges from the noise, which is both interesting and statistically most useful.
This would all be solved if the predictive text engine knew the context beyond what was recently typed. I don't think this is typically true (yet?) -- it would be interesting if all predictions were at least binned by partner vs slang1 friend vs slang2 friend vs business acquaintance et al. with structured data coming to it from the app. Bit of a privacy question I suppose.