Rule scoring framework

This repository contains the set of scripts that has been used for the AMTA paper.

Johann Roturier, Linda Mitchell, Robert Grabowski and Melanie Siegel: Using Automatic Machine Translation Metrics to Analyze the Impact of Source Reformulations, AMTA 2012

The main script is score-rules.sh. The script...

• runs the AutoApply client on a given input text file, creating a set of original and corrected segments in two parallel files, in addition to information about the flag in a third file
• tokenizes and truecases the original and corrected segments using a source language truecaser and tokenizer
• finds the reference translations of the original and corrected segments using a reference file that is parallel to the input file
• translates the original and corrected segments using Moses (either via the Moses server XML-RPC interface, or via the Google Translate API)
• tokenizes and truecases the translated original and corrected segments using a target language truecaser and tokenizer
• tokenizes and truecases the reference translations using a target language truecaser and tokenizer
• scores the tokenized original and corrected segments using a language model for the source language
• scores the translated+tokenized original and corrected segments using a language model for the target language
• scores the translated+tokenized original and corrected segments against the tokenized reference segments using smoothed BLEU, TER, and GTM
• collects all scores and compares them, resulting a statistics CSV file with better/equal/worse rankings grouped by Acrolinx rule

Progress messages are output and written to a log file. The script usually does not create a data file if it already exists.

Usage

Prerequisites:

The script needs the following additional software:

• autoApplyClient, which needs to be requested from Acrolinx
• BLEU scorer
• GTM scorer
• TER scorer

Please look into the README.md files in the respective directories for information on how to obtain and/or compile the software.

Also, the script needs the tokenizer and truecaser scripts from the mosesdecoder repository at https://github.com/moses-smt/mosesdecoder. The root path to your local mosesdecoder repository is hard-coded in the script file as the variable MOSES_DIR at the beginning. Please change that first.

General:

score-rules.sh experiment-name text-file src-lmodel src-tcmodel src-toklang mosesServer ref-file-1 ref-file-2 tgt-lmodel tgt-tcmodel tgt-toklang [autoApplyOptions]

mosesServer:

• either e.g. localhost:8081 (to translate via mosesserver XML-RPC),
• or e.g. http://server/translate.php (to translate via Translate API)

autoApplyOptions:

• e.g. -h host -p port -u user --pass pwd -l lang -r ruleset (run autoApplyClient.jar for complete list)

src-tcmodel and tgt-tcmodel:

• can be "" to skip true-casing

Example:

bash score-rules.sh amta_experiment symc_bip_15_source.clean 1002-10.binlm truecase-model.1.en en 'http://user:pass@accept.statmt.org/demo/translate.php' symc_bip_17_target1.de.clean symc_bip_19_target2.de.clean 1002-06.binlm truecase-model.1.de de -h accept.acrolinx.com -p 80 -u USER --pass PASS -l en -r Preediting_SMT_SYMC -t Symantec -ts DEPRECATED,ADMITTED,VALID -skipReuseCheck

Output files:

• amta_experiment.summary: The complete result, one record per correction instance: autoApply output, Moses translations, and all scores
• amta_experiment.statistics.csv: The summary aggregated/grouped by rule.

A lot of other amta_experiment.* files are written into current directory, so it's best to run the script from a dedicated folder.

Notes

The script still contains a lot of code duplication that should be factored out to make it more flexible.