art: the arbiter to TSDB bridge

Downloads

Overview

art is a simple program using libtsdb that can submit a parsing or generation job to ACE and record the results in a TSDB profile. art can communicate with a single ACE instance natively, or function as a client of the arbiter distributed job controller.

For parsing tasks, you specify a TSDB profile whose input relation contains the utterances to be analyzed. art submits them to ACE and by default records the results in the parse and result relations (and fills out an informational record in the run record. art can also function in forest-based mode. In this case, instead of recording some number of full derivations in the result relation, art records all of the edges from the parse chart that match a root condition or are reachable from another recorded edge, all the way down to the lexical edges and token edges. These go into the edge relation.

For generation tasks, you specify one TSDB profile whose result relation contains MRSes to generate from, and a second TSDB profile into whose result relation the generated realizations are to be written. The MRSes used as generator inputs are selected from the source profile using the preference relation, such as would be found in an annotated treebank. If the preference relation is empty (or missing?), art can instructed to use the first recorded MRS for each item (-p0), second recorded MRS (-p1), etc.

Usage Examples

Preparing a profile from an [incr tsdb()] skeleton:

art-0.1.2$ ./mkprof -s ${LOGONROOT}/lingo/lkb/src/tsdb/skeletons/english/verbmobil/aged aged-profile
9067  bytes	relations
9072  bytes	item
0     bytes	analysis
0     bytes	phenomenon
...
art-0.1.2$

Parsing a profile using a single ACE, recording best-5 derivations:

art-0.1.2$ ./art -a "~/ace-0.9.13/ace -g ~/erg-1111.dat -n 5" aged-profile
reading results for                1	3 results
reading results for                2	1 results
reading results for                3	5 results
...
reading results for               95	3 results
reading results for               96	5 results
art-0.1.2$

Generating with several ACE processors coordinated by arbiter (target profile and gold profile must both already exist):

(terminal 1)$ /path/to/arbiter -e
(terminal 2)$ ./ace -g erg-1111.dat -n 5 -e -m localhost
(terminal 3)$ ./ace -g erg-1111.dat -n 5 -e -m localhost
(terminal 4)$ ./ace -g erg-1111.dat -n 5 -e -m localhost
(terminal 5)$ ./art localhost -e ~/erg-1111/tsdb/gold/csli csli-profile
917 / 1348 items submitted to arbiter
reading results for                1	1 results
reading results for                2	1 results
reading results for                3	2 results
...
reading results for             1343	5 results
reading results for             1347	2 results
(terminal 5)$
Here, the note about 917 / 1348 items means that of the 1348 items in the input profile, 917 of them had MRSes that could be sent off for processing.

To get a summary of commandline options, use art -h.


©2015 Woodley Packard