Step_by_Step_Training.md

Docs - Table of Contents

• Output formats
• Alignment Format
• Nodes and Edges Format
• Data
• Hand Alignments
• Training
• Config File
• Step by Step Training
• Evaluation
• Alignment Evaluation
• Parser Performance
• Development
• Developer's Guide

---

Step by Step Instructions for Training

##1. Preprocessing the data

Download and extract LDC2013E117.tgz into the directory data/LDC2013E117_DEFT_Phase_1_AMR_Annotation_R3. To make the train/dev/test split:

cd scripts/preprocessing
LDC2013E117/make_splits.sh

Then run ./PREPROCESS.sh to tokenize, align, and dependency parse the data.

##2. Training

(To skip this step, which takes about 3-6 hours, download and extract model weights models.tgz into the directory $JAMR_HOME/models.)

cd scripts/training

Extract concept table:

./cmd.conceptTable.train

Concept identification (stage1) training:

./cmd.stage1-weights

Relation identification (stage2) training:

./cmd.stage2-weights

Search for 'Performance on Dev' in stage2-weights.err for early stopping.

##3. Evaluating

Decode test set:

./cmd.test.decode.allstages

or

./cmd.test.decode.stage2only

Evaluate the predictions using smatch:

${JAMR_HOME}/scripts/smatch_v1_0/smatch_modified.py --pr -f ${MODEL_DIR}/test.decode.stage2only ${TEST_FILE}
${JAMR_HOME}/scripts/smatch_v1_0/smatch_modified.py --pr -f ${MODEL_DIR}/test.decode.allstages ${TEST_FILE}