This video comes from the Allen Institute for Artificial Intelligence (AI2) and first appeared on YouTube on May 21, 2017. The speaker addresses the problem of automating the extension of existing natural language processing (NLP) resources.

The guiding philosophy of at least the first part of the talk is that the vast majority of relations can be described by verbs. Think, for example, of the relationship “spouse” and the verb “marry.” Furthermore, verbs have the potential to describe the ways in which relationships change. Think of marry and divorce in regard to the spouse relation. One aim of the work is to find ways to link verbs to existing relations in Carnegie Mellon’s NELL system. Starting with a corpus of subject-verb-object (SVO) triples, each of these can be associated to a typed verb--thus “person” marries “person.” The speaker describes how to build a classifier that predicts the Nell relation associated to a typed verb. The issue of relationship change is addressed through the use of time-stamped corpora, such as the Wikipedia revision pages.

A further case that the speaker addresses is the one where a typed verb does not map to a relation. In this case, starting with a knowledge base with a high coverage of relations, the idea is to form clusters of typed verbs that can become relations. A hierarchy of the types and a notion of closeness of typed verbs, along with constraints determined by the notions of synonym and antonym (here WordNet can be used), serve as the mechanism for the clustering.

In a final section of the talk, the speaker turns to the issue of translation, more specifically to methods that do not rely on parallel texts. As this is the subject of the speaker’s current research, the material is covered in less detail, but it is in a way the part of the talk with the greatest potential.

There are some real issues with the idea of YouTube publication, at least as it appears here. It was hard if not impossible to hear and understand the audience questions. No bibliography appears in the talk, although there are some terse references. The viewer will need to search the web to fill in the details.