Keywords: Micro Knowledge Graph, Technical Support
TL;DR: We present a technique to construct Micro Knowledge Graphs (micrographs) from short technical support web pages that can be used to drive applications such as chatbots, search and QA.
Abstract: Building a Knowledge Graph (KG) from a large corpus of technical support documents such as IBM Technotes poses a challenge of granularity. These documents are relatively short web pages providing information about some specific item or providing solution for one or few related problems. In order to build a rich knowledge base that can be used for question answering or driving multiturn chat-bots, all useful entities and actions along with their relations must be extracted and represented. When all the entities and relations from each of several thousand documents are extracted and added to a knowledge graph, then it becomes very complex and virtually useless as there would be too many instances of entities, actions and relations appearing in many contexts. Therefore, the KG is normally populated with only one or two key entities and symptoms/actions/advises from each page. This makes easier to construct the KG and performing search in it, but it also leads to the loss of other useful knowledge present in these web pages. In this paper, we propose a set of techniques to create a micrograph for each page (in addition to the actual KG for entire domain/document set). The micrograph stores all the entities and actions in a page and also takes advantage from the structure of the page to represent exactly in which part of a particular, these entities and action appeared (eg, in the problem definition, constraints, solution or related information). These micrographs can be invoked for search or lookup when the corresponding web page becomes focus of attention in the application using the main KG. We define schemas for representing structured and unstructured knowledge present in the technical support web pages. Solutions for a technical problem/issue are often presented as procedure of steps in these web pages. We also propose a technique to extract procedures from these webpages and schemas to represent them in the graph.