Attend Copy Parse

Code for paper Attend, Copy, Parse - End-to-end information extraction from documents (https://arxiv.org/abs/1812.07248) by Rasmus Berg Palm, Ole Winther and Florian Laws.

Pre-train parsers

• Put data files in tasks/parsing/data/{amounts,dates}/{train,valid}.tsv following the format in the sample files.
• Modify tasks/parsing/parser.py: set the type variable to train either a dates or amounts parser.
• Execute PYTHONPATH="$PWD" python tasks/parsing/train.py from the root of the repository

Train

• Put data files in tasks/acp/data. One document per file, following the format in the sample file.
• Modify split files in tasks/acp/splits. One document per line
• Modify field in AttendCopyParse to train on different fields. Valid values are [number, order_id, date, total, tla, tta, tp]
• Execute PYTHONPATH="$PWD" python tasks/acp/train.py from the root of the repository

Test

• Modify restore_all_path in tasks/acp/acp.py to the saved model to restore weights from, e.g. ./snapshots/acp/best.
• Execute PYTHONPATH="$PWD" python tasks/acp/test.py from the root of the repository

General

• Every 20 training batches the eval split is evaluated. If the eval loss is better than the best seen so far the model is saved under ./snapshots
• Tensorboard summaries are logged to /tmp/tensorboard

Ideas for future work

In order of difficulty

• Apply to more domains
• Better non-latin support by using better character set (maybe byte-pair encoding)
• Handle multiple pages
• Remove the need for N-grams
• Take field dependencies into account, e.g. total fields should add up.
• Output invoice lines

Screenshots

The scalar values logged

Text output samples logged

Attention distribution logged