In this paper, we introduce \unscripted conversation" - free form dialog over a domain knowledge graph. We describe a use case around Luggage handling for a commercial airline where we answer users queries regarding various policies such as luggage dimensions, restrictions on carry-on items, travel routes etc. We have encoded the domain entities, relationships, processes and polices in the knowledge graph and created a generic semantic natural language processing engine to process user queries and retrieve the correct results from a knowledge graph.
The conventional approach for building chatbots requires explicit conversational
modelling involving manual scripting of dialog ows and extensive training for
intent classi cation. For use cases involving complex organizational processes
and rules like a Luggage Handler, changes in the rules / policies at the backend
necessitates changes in the encoded ows and training data. We believe issues
regarding rigid design and maintenance signal a need for a shift in conversational
systems design. Hence, we look to use Knowledge graphs which are easy to
understand by a human expert, amiable for customization, and easy to maintain.
Knowledge graphs for answering users queries has been used previously [
Our approach centres around the building of a domain graph using a domain agnostic semantic schema and a semantic parsing layer that processes a user utterance and performs a multi-hop graph walk to retrieve the result. 0 Copyright c 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). 2.1
The domain graph follows a at structure where attributes of a entity are the
leaf nodes. Essentially all nodes of the graph inherit from a set of 6 distinct
node types namely Person (for people or roles), Process (for organizational
processes), Math ( for handing math formulas ), Decision ( for if-then-else rules
/ rule expressions ), Generic ( for all other domain entities), Value Property (
for attributes ). In addition, hierarchies de ned in ConceptNet are used to map
their other real world concepts to entities modelled in the graph.[
The semantic natural language layer consists of 4 parts : Query Classi er which classi es the query as info / question and identi es the result type ( a rmation, process, calculation, attribute retrieval, etc ) , the Sentence Splitter which splits a complex sentence into its component clauses, the Base Layer which performs dependency parsing and semantic role labelling, the Custom Layer which builds on previous layers to identify the main clause, focus entities, verbs and the constraints as shown in Figure 2.The entities and verbs are compared with the nodes and relations in the graph using a combination of levenshtein distance and word vectors. The answer is retrieved by nding the shortest path(s) between the start node and the focus node. In scenarios where multiple answers exist (multiple valid identi ed paths / multiple valid focus nodes), a dialog with the user is initiated to retrieve the entities which lead to the selection of a single path. The response is presented to the user via template based natural language generation techniques.