HollyChat! Domain Specific Conversational Agents

Опубликовано 24 июля 2017, 20:22

Most of the AI systems today are driven by three key components (i) data (ii) common sense knowledge and (iii) powerful learning algorithms which can harness this data and knowledge to learn task specific meaningful patterns. Recently there has been a lot of interest in domain-specific dialog systems with several downstream use cases such as shopping assistants, customer support, tour guides, etc. Most of the existing dialog systems are partly in line with the trend mentioned above - in the sense that they are data driven and use powerful algorithms (deep recurrent neural networks and their variants). However, we are nowhere close to building deployable domain-specific conversation systems. One of the primary reasons for this shortfall is that such agents do not exploit any common sense or real-world knowledge, and thereby are unable to maintain a richer context of the conversation. We propose to focus on domain specific conversation systems which use domain specific knowledge graphs as external memory. The idea is to use a domain-specific knowledge graph to discover the latent intent of the user and keep the conversation coherent with this intent. For example, the knowledge graph driven intention network could map the user's utterance \textit{"I really liked the action sequences in Inception (movie)"} to all tuples containing the entity \textit{"Inception"} and keep the conversation restricted to concepts linked to this entity. An appropriate response in this case could be \textit{"Yes, indeed, movies directed by Christopher Nolan are known for their action"} which contains entities and predicates linked to \textit{"Inception"}. This would help in the task of dialog planning and also address the problem of keeping track of large contexts (which would be required for longer conversations containing many turns). In this case, the model could learn to abstract out the context in terms of entities and predicates in the knowledge base and just track these and their immediate neighborhood in the knowledge graph.

See more on this video at microsoft.com/en-us/research/v...