How would you describe your role?
Let’s take the example of when a user speaks to Alana. Alana first has to understand what the user means and then decide the most appropriate response. This involves automatically creating the content from various inputs, such as from a news article, a Wikipedia page, real user interactions with Alana, or human-to-human interactions, before finally generating a response which fits naturally into the dialogue.
Simply put, machine learning involves modelling the patterns of these various inputs, and then using these models to predict and infer user intent, select the appropriate content response and generate it.
Why is this crucial to the development of Alana long-term?
Until very recently, commercial dialogue systems dealt with each of these components, that is understanding user intent, selecting the appropriate content via hand-crafted rules. What we are trying to do with machine learning is to automate the underlying processes. We can then adapt more easily to different customers, user utterances and domains with little human intervention. We can also generate a response that adapts to the style of the conversation.
How might someone using Alana experience this in real terms?
What I help create in Alana is probably the least visible to the end user, while at the same time is the most used feature under the hood! When someone engages in a conversation with Alana, they don’t necessarily know how much content has been retrieved, created from scratch or manipulated, what ended up being discarded as less relevant, and what has been reformulated into natural language.
What are some of the wider developments we can expect to see in Conversational AI? How do you think they will impact the evolution of Alana?
Yet again, we are experiencing an AI revolution. We are able to build better models on vast amounts of data, which five or ten years ago were considered totally science fiction. There are systems out there that can complete whole paragraphs of text or even maintain a long conversation with a user that seems natural. What is yet to be achieved is reasoning in the way we humans do it. Not just extracting facts that follow nicely after what was previously said, but formulating a line of thought: the ability to generate a set of arguments backed by facts. My vision, both as a researcher and as Head of Machine Learning for Alana, is to enable the development of such models.