The Verb Physics project explores how our choice of verbs entails relations between the physical properties of the objects we talk about.

Mary threw _____.

Whatever Mary threw (a ball? a rock?) is probably smaller and weighs less than her

Ricardo walked into _____.

Whatever Ricardo walked into (the library? his office?) is probably larger than him


Explore an interactive visualization of our factor graph model on the Verb Physics dataset. Click and drag on components of the factor graph to move them around.

(Work in progress. Only size-threw_d currently available.)

Type below to select an action frame to visualize. All action frames names start with one of the five attributes: "size," "weight," "strength," "rigidness," or "speed."

Completions (live) (clickable):


The interactive diagram draws a small piece of the factor graph that is focused on the selected action frame. The colors correspond to the model’s decisions about each random variable. Red indicates a decision that a random variable should take the value >, blue represents <, and grey represents =. (Grey is uncommon).

These decisions have different meanings depending on what the random variable represents. There are two different types of random variables:

  1. Object pairs - If a random variable represents two objects—for example, person_vs_house—then the decision for that random variable represents the model’s choice about the relation of those two objects along the given attribute. For example, if we are looking at an action frame for size, then we would expect person_vs_house to take the value <, because people are generally smaller than houses.

  2. Action frames — If a random variable represents an action frame—for example, threw_d—then the decisions for that random variable represents the model’s choice about the relation of two objects that would fit in that action frame. For example, if we are looking at an action frame for size, then we would expect threw_d (which represents <person> threw <object>; see below for more details) to take the value >, because people are generally larger in size than the objects that they throw.

Action frame names

The format for the action frame names is:


The possible attributes are: size, weight, strength, rigidness, speed.

There are five possible action frame constructions. Each corresponds to a syntactic template.

Construction Syntax template Example Example sentence
d <person> <verb> <object> threw_d “I threw the rock.”
od <object1> <verb> <object2> hit_od “The tape hit the ground.”
p <person> <verb> <preposition> <object> threw_p_out “I threw out the trash.”
op <object1> <verb> <preposition> <object2> landed_op_in “The trash landed in the bin.”
dp <person> <verb> <object1> <preposition> <object2> threw_dp_into “I threw the trash into the bin.”


Learning commonsense knowledge from natural language text is nontrivial due to reporting bias: people rarely state the obvious, e.g., “My house is bigger than me.” However, while rarely stated explicitly, this trivial everyday knowledge does influence the way people talk about the world, which provides indirect clues to reason about the world. For example, a statement like, “Tyler entered his house” implies that his house is bigger than Tyler.

In this paper, we present an approach to infer relative physical knowledge of actions and objects along five dimensions (e.g., size, weight, and strength) from unstructured natural language text. We frame knowledge acquisition as joint inference over two closely related problems: learning (1) relative physical knowledge of object pairs and (2) physical implications of actions when applied to those object pairs. Empirical results demonstrate that it is possible to extract knowledge of actions and objects from language and that joint inference over different types of knowledge improves performance.


A picture of Maxwell Forbes

Maxwell Forbes

A picture of Yejin Choi

Yejin Choi


The paper is available on arXiv.

a thumbnail rendering of the ACL 2017 verb physics paper


  title = {Verb Physics: Relative Physical Knowledge of Actions and Objects},
  author = {Maxwell Forbes and Yejin Choi},
  booktitle = {ACL},
  year = {2017}


The data is available in the verbphysics GitHub repository under data/.

See the repository README for more information on the data splits and task setup.


Visit the verbphysics GitHub repository for our reference implementation and instructions for running our code.

It is released under the permissive MIT license.