285 — KG²: Learning to Reason Science Exam Questions with Contextual Knowledge Graph Embeddings

The AI2 Reasoning Challenge (ARC) is essentially a science exam for machines: It emulates a basic test-taking format, and is used as a benchmark for the question-answering artificial intelligences of THE FUTURE

One example question:

Which property of air does a barometer measure? (A) speed (B) pressure (C) humidity (D) temperature

You can answer this question pretty easily by using a few naive methods: “Information Retrieval” (IR) is basically looking up the answer in a dictionary; “word-co-occurrence” checks a large natural language corpus and determines which of the phrases most closely co-occurs with the question. (Terms like “property” probably cooccur equally often with all of the answers, but terms like “barometer” and “pressure” probably pop up close to one another in the training data comparatively more often.)

But not all questions can be answered this way — especially if a unique vocabulary, or a term outside of the IR’s domain, is used.

The authors of this paper propose a new technique to answer these sorts of questions: The technique is a marriage of knowledge graphs (#knowledge-graph) and machine learning (#machine-learning) that “learns” a graph embedding (see also #104) in order to improve graph traversal when seeking an answer.

This sort of technology is likely very useful in fields such as medicine, where a large text corpus (e.g. textbooks, web resource, etc) needs to be interrogated for complex questions (e.g. diagnoses).