Tags

New intent generation + perf and cosmetic improvements

No more parsing errors, good for all 1088 pages

Better follow RASA training file conventions

The json produced before did allow to train a model, but examples are
given in a specific style. Automated tools use it (rasa-nlu-trainer
specifically), and in doing so, their changes will be easier to keep
track of.

Switch `text` and `intent` positions in the examples, and added an
empty list of `entities` at the end.

Restrict examples to text-intent pairs (see #11)

Select columns `text` and `intent` in the transform method to avoid creating
unnecessarily big json objects.

Separate modules to download and process data

- `download_raw_data` downloads the webpages
- `etl` parses the html content, creates a json file from the selected data
and outputs in in the specified directory

See #8.

Smarter crawler, full parser and generate json

- The crawler now uses the "hidden menu" found in a js script from any
page, follows all links.
- The parser extracts the question, answer, category, tags and related
questions.
- A json file conforming to rasa training data format is generated from
all questions. The `intent` of exemples are generated from the
`question` field.

Naive crawler: generate qid 1-2000