ALOOF Relations

In this pipeline, we implemented some methods to extract relations (such as hasShape, isA, partOf, among others) about objects that are usually located in domestic evironments (house, rooms). This implementation is part of the project ALOOF (https://project.inria.fr/aloof/).

The goal of ALOOF is to equip autonomous systems with the ability to learn the meaning of objects, i.e. their perceptual and semantic properties and functionalities, from externalized knowledge sources accessible through the Web.

To run this pipeline, it is necessary to have installed Python 2.7.

Installation

After clone the repository or download the source code, you must install the prerequisite Python packages listed in the file requirements.txt.

With pip, this is done with:

$ pip install -r requirements.txt

Configuration

In src folder you can find the code and in resource folder the resources that we were used.

If you want to change the paths of the input data, in the main.py file you must to change these options:

house_objects_path = "YOUR PATH"
frame_raw_path = "YOUR PATH"
visualgenome_raw_path = "YOUR PATH"
conceptnet_raw_path = "YOUR PATH"

Execution

The general way to run this program is the following:

 $ cd src/
 $ python main.py option

Where option could be:

"objects", to extract house's objects.
"attributes", to extract attributes of objects.
"frames", to extract relations using Frame Semantics.
"conceptnet", to extract relations from ConceptNet.

Example

 $ python main.py attributes

Outputs

This pipeline generates RDF triples in nt format (line-based).

File selected_triples.nt:

<http://dbpedia.org/resource/Spoon> <http://ns.inria.fr/deko/ontology/deko.owl#usedFor> <http://dbpedia.org/resource/Eating>
<http://dbpedia.org/resource/Spoon> <http://ns.inria.fr/deko/ontology/deko.owl#locatedAt> <http://dbpedia.org/resource/Dishwasher>
<http://dbpedia.org/resource/Spoon> <http://ns.inria.fr/deko/ontology/deko.owl#isA> <http://dbpedia.org/resource/Utensil>

File selected_triples_label.nt:

<spoon> <usedFor> <eating>
<spoon> <locatedAt> <dishwasher>
<spoon> <isA> <utensil>

Notes

For the extraction of attributes, we used the model Word2vec. Specifically, we used the pre-trained vectors obtained from part of Google News dataset (about 100 billion words). This file is large and is not in this repository. You can download the file here and use it to replace the file resource/embeddings/googlenews_negative300.bin in this repository.

damengde / aloof