Improve per-word sentiment analysis

Lots of words seem "off" (e.g. phone, surf being negative). Need to investigate a way to make these better than just a flat list of word => scores, since different regions are going to feel differently about words and manually adjusting word scores does not scale.

Should probably also use chunking for sentiment, instead of per-word checking.

This is probably the most common feedback received so far.

Ideas?

No files, tests, or entry points are named. Start by locating the current flat word-to-score sentiment implementation and reproduce the reported results for “phone” and “surf”; then investigate regional variation and chunk-based sentiment before defining a testable scope and completion criteria.

Written by the indexing model from the issue text.