Link to paper This is the seminal paper on a linguistic theory called centering, which deals with “relationships among focus of attention, choice of referring expression, and perceived coherence of utterances within a discourse segment.” In other words, this theory will tell you why “John has been acting odd. He called up Mike yesterday. He was annoyed by John’s call.” sounds a pretty funny, and what the rules are that govern (or influence) our resolution of pronouns.

Semantic role labeling is the task of determining what the arguments of a verb are. For instance, in “The lion ate the deer,” the verb “ate” has two semantic roles: the agent (the lion) who does the eating, and the patient (the deer) who was eaten. The roles stay the same even if the syntax changes; in “the deer was eaten by the lion,” the deer is now the syntactic subject of “was eaten,” but it is still the patient in terms of semantic roles.

This is Watson, a series of papers describing IBM’s Watson system that beat Ken Jennings at Jeopardy.

After a long hiatus, I think I will start posting on research papers again. I’ve been reading still, I just have been pretty busy with other stuff so I haven’t written about them. Here goes.

The premise of this paper is that current distributional measures of word similarity are deficient. Topic models and context vectors only give a single measure, and intuitive measures of similarity using these models do not satisfy the triangle inequality (“bat” is similar to “club,” and “club” is similar to “association,” but the sum of those two distances is less than the distance from “bat” to “association”). Also, measuring similarity between each pair of words might require using different features for each pair (“wine” and “vinegar” are similar, as are “wine” and “bottle,” but each pair is similar for different reasons).