Spatial Language Understanding Framework
The Spatial Language Understanding Framework consists of several
components: corpora and python software to read the datasets, a
spatial feature library, trained models for individual words, and an
end-to-end language understanding system.
All the code is released in the
PODS format for ease
of integrating with larger projects.
SLU Core
This package contains the complete command understanding system,
including the supervised models for learning word meanings. It also
contains many of our datasets, described below.
$ git clone http://github.com/h2r/slu_core
Corpora
The format for our annotated data is a custom
YAML format which contains information about
the groundings (the context), the language and parse tree, as well as
alignments between the language and the groundings. We provide
the data files as well as python code for reading the datasets. The
datasets are divided by paper.
- aaai11 Stefanie Tellex, Thomas
Kollar, Steven Dickerson, Matthew R. Walter, Ashis Gopal Banerjee,
Seth Teller, Nicholas Roy. (2011). Understanding Natural Language
Commands for Robotic Navigation and Mobile Manipulation. Proceedings
of the National Conference on Artificial Intelligence (AAAI), San
Francisco, CA.
- Recipes
- Unaligned Parallel Data
- Spatial Relations
SLU Features
The SLU features library is a standalone library that consists of two-
and three- dimensional feature vectores we use to compute grounded
word meanings. It is bundled with slu_core, so you only need to
download it separately if you wish to use it outside of slu_core.
There are several families of features depending on the type of the
grounding (e.g., path, prism, and whether it is 2d or 3d.)
You can download the library from the git repository:
$ git clone http://github.com/slu/slu_features
The README in that directory has instructions on how to use it.