Zana Health Assistant now leveraged with Co-reference Resolution within Zana Brain
Our team at Zana has further improved the capabilities of the intelligent health companion to understand the various ways users express their health concerns.
Zana is an interactive artificial intelligence health assistant that can respond to questions with accurate answers, concise informative articles, and personalized recommendations.
As we know, natural language is full of ambiguity, reference, and context — challenges that we are addressing in our everyday developments to improve the assistant. One of the challenges we have recently tackled is the co-reference anaphora resolution problem.
What is co-reference resolution?
Co-reference in conversational agents consists of finding entities in the current user request that refer to a previous entity, which we refer to as key concept. Let’s try to understand this by having a look at the following conversation with our assistant Zana.
In this conversation, one can notice that Zana recognized the in the user`s utterance related to the medical condition anaphylaxis, and understood the intent of the user to ‘ask for introduction’ of such condition.
In the next utterance, the user then implicitly asks about the treatment of anaphylaxis with the pronoun “it”. The ability to link such reference in natural language representation is called co-reference/anaphora resolution.
We extended Zana understanding capabilities by building and integrating a co-reference resolution module as part of our ‘Zana Brain’ — the engine that resides in the core of our technology.
What makes this setting more interesting, and at the same time challenging, is that we deal with the co-reference problem in a conversation flow which is much more difficult than resolving the reference only one particular request (just one string). We additionally need to keep track of the contextual information in the conversation.
Our approach follows these two steps:
1. Based on our existing NLU pipeline we are able to detect the user intention, to recognize the appropriate medical-related entity that serves as the key concept in the dialogue flow. This key concept relevant for the resolution of the anaphora-antecedent problem.
2. The second step is to recognize and classify those requests that refer to the previous key concept. The new implementation performs the co-reference resolution in order to link the current user intention ‘ask treatment’ with the key concept recording in the conversation e.g. in this case anaphylaxis.
Zana is now able of capturing misleading queries that refer to a medical topic mentioned and captured as a key concept previously in the conversation. This allows the user to have a more human-like conversation with our AI-powered assistant without the need to repeat certain entities and to make redundant questions on and on.
In the conversation below, you can see how Zana understands the reference and sends back to the user the implicitly asked information about the medical topic.
A s a team, we are working on the most amazing and challenging subsets of natural language tasks in order to enrich Zana intelligence.
We will follow with other updates very soon!
