They suggest that they add the word vectors in the sentence. But it seems to me that that would make the result independent of the order of words (i.e. "when does Tesco open?" and "Open Tesco when does" are the same). I thought I had tested that and it didn't work but actually I just tried saying "Tesco open does when?" to Alexa and it said "Sorry, I don't have the business hours for Tesco". Inconclusive I'd say but interesting anyway!
yeah you're quite right, intents are built with a bag of words model & doesn't take order into account. Entity extraction does though. If you find a case where word order is really important for getting intents right I'd love to know about it! We could find a way to make that work.
The bot I have been working on uses bag of words and n-grams to identify intents. Useful for when someone says something that may have multiple entities of the same type. For instance a journey start and end point. Can use the "from" and "to" words match better. Also very useful when training on phrases that are very similar such as FAQ questions which often have the same words but the order and one or two keys words is super important to get the right match.
It's the "conversation" part that is really tricky... I have been working on a bot for a large Corp for the last few months and we have been using Inkle's Ink narration/dialog engine for this. Works very well. They let me open source the framework: https://github.com/rabidgremlin/Mutters it uses OpenNLP for intent identification and NER, Ink for conversation state and "scripting"
This is interesting, I've been using LUIS for some time now and an open source alternative - especially one that is drop-in API compatible - is very welcome.
However I can't find any information in the docs on how comparable the results are (i.e. does it have built-in date and time entity recognition like LUIS?). Most importantly: what languages does this support? All examples are in english-only. Is it even language aware, or do you train a model in any language? I'd be very interested if this were to support languages that LUIS does not have (like my language: Dutch)
Currently it supports english and german. In general we need a word embedding for each language. If that has been created by someone else, it's rather easy to integrate new languages.
currently there are no built-in entities like dates, times, locations etc. But really keen to set up a way for users to share models, and that would definitely include these things as well
Looks like an interesting project, based on skilearn and spaCy. The project provides some simple training files for the domain of asking about restaurants.
It would be useful to also have very large training data sets available.
That said, from my experience, you can get surprisingly far with simple systems; for instance, queread[2] relies on graph learning and statistics.
[0]: https://github.com/mit-nlp/MITIE
[1]: https://spacy.io/
[2]: https://github.com/espadrine/queread#workings