Running BIUTEE

Table of Contents BIUTEE Pipeline Running BIUTEE from the Shell Pre-trained Models More Options

BIUTEE Pipeline

BIUTEE runs in three main phases:

1. Preprocessing (LAP)
2. Training
3. Testing

Given two XML files in RTE format, one for train and one for test, each needs to be preprocessed separately before training/testing. This converts the original text into an annotated dependency tree, and stores it in an internal format. This is useful, since one may want to experiment with different configurations of BIUTEE without running the entire LAP over and over again.

BIUTEE can then train a model from the preprocessed training data. This model is then used by BIUTEE during inference (the testing phase).

Running BIUTEE from the Shell

In Linux/Unix, BIUTEE can be run from the command line using run_biutee.sh from $BIUTEE/workdir. Running with -h reveals the modes and options:

 run_biutee.sh [OPTIONS] MODE
 
 Available modes:  lap_train
                   train
                   lap_train,train
                   lap_test
                   test
                   lap_test,test
                   visual
 
 -i <filename>     raw input file (rte pairs format)
 -s <filename>     serialized input file (saves/loads the preprocessed pairs after LAP)
 -m <dir>          models directory
 -o <filename>     results file (only in test modes)
 -r <list>         list of resources, separated by commas (no spaces)
 -t <int>          number of threads
 -e <port>         easyfirst parser's port
 -f                optimize F1 (instead of accuracy)
 -g                gap mode

The different modes are thoroughly explained in the following section. For first-time use, we recommend running lap_train,train followed by lap_test,test.

Pre-trained Models

BIUTEE comes along with a few pre-trained models. They are all located under $BIUTEE/workdir/models/. The following table describes how each one was trained.

More Options

Detailed Guide on Running BIUTEE

Running BIUTEE's visual tracing tool