This week, I made some progress regarding the problem of matching users’ spoken requests to the relevant responses.
Rasa – a machine learning framework
When researching SpaCy, I looked into some of the projects that used it internally. One of those was Rasa, an open-source machine learning framework for automated text-based conversations.
It has a natural language understanding (NLU) model that can be used to build conversational assistants, even with a fairly limited sample of training data. It seems relatively flexible, so it should be possible to integrate it within our existing AskBob code.
Rasa is written in Python and allows us to use SpaCy components as a part of the NLU pipeline to take advantage of SpaCy’s powerful feature-set (including its entity extractor). It also means we can use it to build chatbots in any of the languages supported by SpaCy.
Rasa also supports other ‘connectors’, such as Facebook Messenger and Slack in addition to a REST API, which could be useful if we ever wanted to integrate AskBob into other platforms.
Client meeting
We also held a client meeting with Dr Dean Mohamedally where we took stock of the term’s progress and considered some of the things we might want to focus on once we have a prototype voice assistant, including the following:
- testing AskBob on low-power devices, e.g. the Raspberry Pi
- building plugins to integrate third-party APIs into AskBob
- researching
Voiceflowto either use it or build a similar intuitive interface to design voice commands - collaborating with another systems engineering team building a concierge app
Next steps
Over the Christmas break, it would be useful to explore Rasa’s documentation and try out a few examples to learn more about how we could integrate it into our project in order to build a prototype voice assistant.