DIRE Dataset

Corpus

This page provides the dataset from Boleda et al. IWCS 2017. The dataset consists of six individual files:

• stimuli.train.gz, stimuli.valid.gz, stimuli.test.gz: The stimuli themselves, one sequence per line, for train set 
  (40K sequences), dev set (5K sequences), and test set (10K sequences). Total size: 4.5 MB.
• image.dm.gz: The corresponding image vectors (from Lazaridou et al. NAACL 2015). Größe: 167MB.
• word.dm.gz: The corresponding word embeddings (aus Baroni et al. ACL 2014). Größe: 2.5MB.

The syntax of the stimulus files is as follows:

line      = query query_position || entities || stimuli
query     = category:modifier:modifier
entities  = 6(entity )  
entity    = category_picindex
stimuli   = 12(modifier:entity )

The values of "category" serve as keys in word.dm, and the values of "entity" as keys in image.dm.gz.
These two files are simple line-based hash tables with the syntax "key value" which map string keys onto vectors.

The DIRE implementation is available on this page: TBC.

Living a discrete life in a continuous world: Reference in cross-modal entity tracking.
Proceedings of IWCS. Montpellier, France, 2017.
Gemma Boleda, Sebastian Padó, Nghia The Pham and Marco Baroni.