cjvt-srl-tagging/tools/srl-20131216/scripts
2019-01-25 07:29:52 +01:00
..
learn.sh added mate-tools srl parser 2019-01-25 07:29:52 +01:00
parse_full.sh added mate-tools srl parser 2019-01-25 07:29:52 +01:00
parse_srl_only.sh added mate-tools srl parser 2019-01-25 07:29:52 +01:00
README added mate-tools srl parser 2019-01-25 07:29:52 +01:00
run_anna_http_server.sh added mate-tools srl parser 2019-01-25 07:29:52 +01:00
run_http_server.sh added mate-tools srl parser 2019-01-25 07:29:52 +01:00

This folder contains some scripts to execute the system. They all need
to be edited slightly before use with proper paths to corpora and/or
models. They are meant to be executed from the parent directory, e.g.

  cd $DIST_ROOT
  sh scripts/learn.sh

There are some comments in the script on switches etc. The amount of
memory used is typically whats required for the English CoNLL 2009
corpora. It might be possible to push it down a bit.


Since the system grew out of the CoNLL 2009 ST, there are a couple of
different ways to parse a corpus:

(i) parse_full.sh - parses a complete corpus using all steps of the
                    pipeline except tokenization. It pretty much
                    assumes that that the file contains tokens in the
                    second column and disregards the rest. 
                    If the -nopi switch is used, it needs to have the
                    IsPred column from the CoNLL 2009 data format.

(ii) parse_srl_only.sh - parses semantic roles only. The input is
                         expected to be the CoNLL 2009 data format
                         with proper dependency trees (ie. the
                         SRLonly evaluation corpus).
                         In order to replicate the setting of the 2009
                         ST, one can use the -nopi switch to skip the
                         predicate identification step.


Then there is also the HTTP interface. This is started by the
run_http_server.sh script. Again, edit the file with proper paths
before executing it.

NOTE: the http server depends on the java package
com.sun.net.httpserver (cf.
http://download.oracle.com/javase/6/docs/jre/api/net/httpserver
/spec/com/sun/net/httpserver/package-summary.html
and
http://blogs.sun.com/michaelmcm/entry/http_server_api_in_java
), which is not part of the real Java specification, but comes with
most (or at least some) JRE distributions. From my own experience,
it is included in the Sun Java 6 distribution* as well as the OpenJDK 
Java 6**.

[[
 *: 
  On a Mac:
  % java -version
  java version "1.6.0_17"
  Java(TM) SE Runtime Environment (build 1.6.0_17-b04-248-10M3025)
  Java HotSpot(TM) 64-Bit Server VM (build 14.3-b01-101, mixed mode)
  %

 and
  On pc, mandriva linux:
  % /usr/lib/jvm/java-sun/bin/java -version
  java version "1.6.0_15"
  Java(TM) SE Runtime Environment (build 1.6.0_15-b03)
  Java HotSpot(TM) 64-Bit Server VM (build 14.1-b02, mixed mode)
  %

 **:
  On pc, mandriva linux:
  % java -version
  java version "1.6.0_18"
  OpenJDK Runtime Environment (IcedTea6 1.8) (mandriva-2.b18.2mdv2009.1-x86_64)
  OpenJDK 64-Bit Server VM (build 14.0-b16, mixed mode)
  %
]]

The graphical dependency graph output of the HTTP interface relies on 
having Chinese fonts install properly locally. On Linux I had some issues
with this, but resolved them according to
http://blog.lizhao.net/2007/03/java-chinese-fonts-on-ubuntu.html

The graphcical dependency graph output also seems to work less good 
when using OpenJDK. The images get strange lines through them. The Sun
JRE seems to work fine though. 


Feedback and questions are appreciated: anders@ims.uni-stuttgart.de