IBL CORPUS

This corpus has been collected as part of the IBL project funded by the UK Engineering and Physical Sciences Research Council and jointly managed by:

Dr. Guido Bugmann at the Centre for Neural and Adaptive Systems of the University of Plymouth (GR/M90023).
Dr.
Ewan Klein the Institute for Communicating and Collaborative Systems (ICCS) of the University of Edinburgh (GR/M90160).

The IBL project aims at exploring methodologies for creating new programs in domestic or helper robots through verbal instructions using unconstrained natural language. For more informations about the project see: http://www.tech.plym.ac.uk/soc/staff/guidbugm/ibl/index.html
 

Information on the Corpus:

- map used
-
Procedure Description
-
Audio and Transcribed files details
-
Accessing the files


Map used:
 


 

Letters on top of the buildings label the various building. There is a lake and a bridge.
Real dimensions: 168 cm x 119 cm ( four A0 pages).

List of building:



Procedure Description:

The aim of this first series of tests is to extract expressions and words used by humans in these conditions (i.e. instructing an agent moving in an artificial environment) in order to define a lexicon. The second aim is to fix the level of complexity for the robot in-built primitives (i.e. the primitives the robot can execute without any training).

The main features for this test are:

It is planned to use 24 participants for the test, where each participant has to give 6 route instructions.
To satisfy the third point in the 'main features' list,  the following criteria have been adopted when the various route set have been defined (see table 1). 6 different routes are given to each user. They will be coupled in pair, so that for each route there are 3 subset, each containing a long route L and a short route S, where S is a subpath of L (for example M-C=L and M-W=S is the first subset for route M) .
For each participant the 6 different routes start all from the same place. In total there will be 4 different starting point. That is, 4 different participants will use the same route set . In this way it should be possible to analyse the correlation between the sequence order and the way participants combine different paths.
(see the map for more details about the associated list of building) :  


Audio  and Transcribed Files Details

AUDIO FILES
In the corpus there are 146 files. Each audio file in each directory has a name like:
u5_GC_EH.wav

where the first two digits indicates the user# (ie u1=user1) the second two digits gives the GROUP test (either A, B,C) (GA=Group A(monologues), GB=Group B (monlogues),GC=Group C (dialogues)), while the last couple of digits before the dot (.) give the starting point and ending point (for example starting in E ending in H (from the Grand Hotel to the Hospital) see table below). Finally the digits after the dot give the file format (wav= wav audio format)
NOTE: there are also 2 extra files included as *bis.wav.
 

The table below shows a summary of the possible routes.
 

ROUTE SUMMARY TABLE.

Route References Route M (Short/Long) Route E (Short/Long) Route C  (Short/Long) Route H (Short/Long)
M Museum

W University

X Post Office

C Boots

Y Library

E Grand Hotel

D Safeway

L Queen's Pub

G Tesco

P Car Park

H Hospital 

M-W S E-X S C-D S H-L S
M-C L E-G L C-E L H-G L
M-X S E-W S C-P S H-D S
M-Y L E-C L C-M L H-E L
M-E L E-P L C-L L H-C L
M-D S E-H S C-X S H-W S

For each Route * column (Route M, Route E, Route C, Route H), there are 2 sub-columns. In the first one each letter in the pair (*-* ) indicates respectively the starting and ending poin in the route used. In the second column the letter (S or L) indicates whether the route is a Short (S) or Long (L) ones.  The Route References gives the building (as indicated in the aerial picture of the maze) associated to each letter.
For example for Route M then, the SLSLLS sequence corresponds to M-W,M-C, M-X,M-Y, M-E, M-D, while the sequence LSLSSL is given by M-C,M-W,M-Y,M-X,M-D,M-E.
 


Accessing the files

The recorded instructions are chunked into smaller sound files with one utterance per file (chunking is based on silences). 
These files (approx 60MB) can be downloaded as a
gzipped tarfile containing the files in NIST-wav format. See players and conversion tool.
There is also a text file with
transcriptions of the utterances and
a set of functional transcript files
Funct_Transcript_training_set.zip and Funct_Transcript_test_set.zip that contain the sequence of robot function calls corresponding to the verbal instructions.

Please contact either Guido Bugmann or Ewan Klein for copies on CD or for the original unchunked files (see links at the top of this page).

LICENSE.
Everyone is permitted to copy and distribute copies of the IBL CORPUS for research or educational purpose only.
This License applies to any audio and text files described in this document (IBL CORPUS) which the members of this project make publicly available.


Last updated 24/6/2004