SUDA-HUAWEI at CAMRP2022

This repository contains the system we submitted at CAMRP2022. Our model is built based on PERIN. Report is supported here.

Data Preprocess

From raw file to mrp file.

# first use preprocess/filte_duplicate.py to filte duplicate notes
python filte_duplicate.py --raw_file camr_dev.txt --out new_camr_dev.txt

# next use mtool to create file in .mrp
cd mtool
./main.py --read camr --write mrp new_camr_dev.txt camr_dev_fdomain.mrp

# here may happen some errors, for example:
# failed to parse #export_amr.1272 ‘(x56 / causation :arg1 (x58 / and :op1 (x1 / 有-03 :arg1 (x4 / 落款-01 :arg0 (x101 / government-organization :name (x2_x3 / name :op1 __0__ :op2 __1__ )))) :op2 (x7 / 有-03 :arg1 (x17 / 要求-04 :arg1 (x16 / 数量) :arg0-of (x68 / mean :arg1 (x12 / 刷-01 :arg0 (x69 / and :op1 (x9 / 乡镇 :mod (x8 / 各)) :op2 (x11 / 部门 :mod (x10 / 各))) :arg1 (x14 / 票 :quant (x13 / 多少)))))) :op3 (x24 / 发布-01 :location (x104 / publication :name (x22_x23_1 / name :op1 __2__ :op2 x23_1/网 )))) :arg2 (x82 / and :op1 (x31 / 能-01 :mod (x30 / 还) :arg0 (x32 / 说-01 :arg1 (x42 / 行为 :poss (x38 / 好心人 :quant (x36 / 个别)) :domain (x33 / 这) :mod (x34 / 只) :arg0-of (x41 / 个人-01))) :mode (x43_x44 / interrogative)) :op2 (x46 / 能-01 :mod (x45 / 还) :arg0 (x47 / 说-01 :arg1 (x51_2 / 有-01 :polarity (x51_1 / -) :arg1 (x52 / 责任) :arg0 (x48 / 你 :arg0-of (x97 / mean :arg1 (x49_x50 / x101))))) :mode (x53_x54 / interrogative))))’; exit.

# This is caused by some "op", for example "op2 x23_1/网", we change these ops to "op2 x23/网"

# next add functional word to edge label
python preprocess/add_fw.py --tup_train ccl2022/camr_official/camr_tuples/tuples_train.txt \
                            --raw_train ccl2022/camr_official/camr/camr_train.txt \
                            --tgt_tup_file ccl2022/camr_official/camr_tuples/tuples_dev.txt \
                            --tgt_raw_file ccl2022/camr_official/camr/camr_dev.txt \
                            --tgt_mrp_file ccl2022/tmp_camr_dev_fdomain.mrp \
                            --out ccl2022/addfw_camr_dev_fdomain.mrp

# add pos and syn
python preprocess/add_pos-syn.py --mrp ccl2022/addfw_camr_dev_fdomain.mrp \
                                 --conllu ccl2022/camr_official/camr_dev.txt.out.conllu \
                                 --out ccl2022/addfw_mytokenize_camr_dev_fdomain.mrp

Install

The installation is mostly the same as Perin, except the mtool. We modified the label normalization to fit chinese amr. The modified mtool is here.

Train

CUDA_VISIBLE_DEVICES=3 python train.py --config config/addfw_chinese_rbt_large.yaml \
                                        --data_directory ccl2022 \
                                        --save_checkpoints \
                                        --workers 1

Inference

CUDA_VISIBLE_DEVICES=5 python inference.py --checkpoint outputs/08-25-22_17-21-05/best_checkpoint.h5  --data_directory ccl2022

Multi-graph Ensembling

python graph_ensemble_aug.py --dir wisyn_testb_graph_ens \
                             --len_file max_len_testB.txt \
                             --outname testb.tup

"--dir" means the directory containg all predictions (in mrp format) of different model
"--len_file" is here.

Evaluation

We use AlignSmatch as the metric. We use the script from camr2022.

zsLin177 / camr