release

Author: franklear

bash/amr_preprocess.sh

@@ -0,0 +1,26 @@
+#!/bin/bash
+
+CURRENT_PATH=$(cd $(dirname "${BASH_SOURCE[0]}") && pwd)
+
+echo '######## converting mrp to amr ...' >&2
+python3 ${CURRENT_PATH}/../utils/mrp_to_amr.py -i $1 -o $1.amr.txt --not_amr_str_only --all_nodes
+
+export JAMR_HOME=path/to/jamr
+export CDEC=path/to/cdec
+
+echo '######## running jamr rule based aligner ...' >&2
+${JAMR_HOME}/scripts/ALIGN.sh < $1.amr.txt > $1.jalign.txt
+
+TOOLKIT_HOME=${CURRENT_PATH}/../toolkit
+
+echo '######## running tamr rule based aligner ...' >&2
+python3 ${TOOLKIT_HOME}/tamr_aligner/rule_base_align.py -verbose -data $1.jalign.txt -output $1.alignment.txt -wordvec $2 -trials 10000 -improve_perfect -morpho_match -semantic_match
+
+echo '######## refreshing alignments ...' >&2
+python3 ${TOOLKIT_HOME}/tamr_aligner/refresh_alignments.py -lexicon $1.alignment.txt -data $1.jalign.txt > $1.new_aligned.txt
+
+echo '######## generating oracles ...' >&2
+python3 ${TOOLKIT_HOME}/tamr_aligner/eager_oracle.py -mod dump -aligned $1.new_aligned.txt > $1.actions.txt
+
+echo '######## adding extra mrp information ...' >&2
+python3 ${TOOLKIT_HOME}/amr_add_extra.py -i $1.actions.txt -o $1.actions.aug.txt -e mrp $1

bash/predict.sh

@@ -0,0 +1,78 @@
+#!/bin/bash
+
+# examples of predicting commands
+
+# DM
+CUDA_VISIBLE_DEVICES=0 \
+allennlp predict \
+--cuda-device 0 \
+--output-file dm-output.mrp \
+--predictor transition_predictor_sdp \
+--include-package utils \
+--include-package modules \
+--use-dataset-reader \
+--batch-size 32 \
+--silent \
+checkpoints/dm_bert \
+data/dm-test.mrp
+
+
+# PSD
+CUDA_VISIBLE_DEVICES=0 \
+allennlp predict \
+--cuda-device 0 \
+--output-file psd-output.mrp \
+--predictor transition_predictor_sdp \
+--include-package utils \
+--include-package modules \
+--use-dataset-reader \
+--batch-size 32 \
+--silent \
+checkpoints/psd_bert \
+data/psd-test.mrp
+
+
+# EDS
+CUDA_VISIBLE_DEVICES=0 \
+allennlp predict \
+--cuda-device 0 \
+--output-file eds-output.mrp \
+--predictor transition_predictor_eds \
+--include-package utils \
+--include-package modules \
+--use-dataset-reader \
+--batch-size 32 \
+--silent \
+checkpoints/eds_bert \
+data/eds-test.mrp
+
+
+# UCCA
+CUDA_VISIBLE_DEVICES=0 \
+allennlp predict \
+--cuda-device 0 \
+--output-file ucca-output.mrp \
+--predictor transition_predictor_ucca \
+--include-package utils \
+--include-package modules \
+--use-dataset-reader \
+--batch-size 32 \
+--silent \
+checkpoints/ucca_bert \
+data/ucca-test.mrp
+
+
+# AMR
+# !!! AMR parser accepts input of augmented amr format instead of mrp format !!!
+CUDA_VISIBLE_DEVICES=0 \
+allennlp predict \
+--cuda-device 0 \
+--output-file amr-output.mrp \
+--predictor transition_amr_predictor \
+--include-package utils \
+--include-package modules \
+--use-dataset-reader \
+--batch-size 32 \
+--silent \
+checkpoints/amr_bert \
+data/amr-test.txt

bash/train.sh

@@ -0,0 +1,83 @@
+#!/bin/bash
+
+# examples of training commands
+
+# DM
+CUDA_VISIBLE_DEVICES=0 \
+TRAIN_PATH=data/dm-train.mrp \
+DEV_PATH=data/dm-dev.mrp \
+BERT_PATH=bert/wwm_cased_L-24_H-1024_A-16 \
+WORD_DIM=1024 \
+LOWER_CASE=FALSE \
+BATCH_SIZE=4 \
+allennlp train \
+-s checkpoints/dm_bert \
+--include-package utils \
+--include-package modules \
+--file-friendly-logging \
+config/transition_bert_sdp_conll.jsonnet
+
+
+# PSD
+CUDA_VISIBLE_DEVICES=0 \
+TRAIN_PATH=data/psd-train.mrp \
+DEV_PATH=data/psd-dev.mrp \
+BERT_PATH=bert/wwm_cased_L-24_H-1024_A-16 \
+WORD_DIM=1024 \
+LOWER_CASE=FALSE \
+BATCH_SIZE=4 \
+allennlp train \
+-s checkpoints/psd_bert \
+--include-package utils \
+--include-package modules \
+--file-friendly-logging \
+config/transition_bert_sdp_conll.jsonnet
+
+
+# EDS
+CUDA_VISIBLE_DEVICES=0 \
+TRAIN_PATH=data/eds-train.mrp \
+DEV_PATH=data/eds-dev.mrp \
+BERT_PATH=bert/wwm_cased_L-24_H-1024_A-16 \
+WORD_DIM=1024 \
+LOWER_CASE=FALSE \
+BATCH_SIZE=4 \
+allennlp train \
+-s checkpoints/eds_bert \
+--include-package utils \
+--include-package modules \
+--file-friendly-logging \
+config/transition_bert_eds.jsonnet
+
+
+# UCCA
+CUDA_VISIBLE_DEVICES=0 \
+TRAIN_PATH=data/ucca-train.mrp \
+DEV_PATH=data/ucca-dev.mrp \
+BERT_PATH=bert/wwm_cased_L-24_H-1024_A-16 \
+WORD_DIM=1024 \
+LOWER_CASE=FALSE \
+BATCH_SIZE=4 \
+allennlp train \
+-s checkpoints/ucca_bert \
+--include-package utils \
+--include-package modules \
+--file-friendly-logging \
+config/transition_bert_ucca.jsonnet
+
+
+# AMR
+# !!! AMR parser accepts input of augmented amr format instead of mrp format !!!
+CUDA_VISIBLE_DEVICES=0 \
+TRAIN_PATH=data/amr-train.mrp.actions.aug.txt \
+DEV_PATH=data/amr-dev.mrp.actions.aug.txt \
+BERT_PATH=bert/wwm_cased_L-24_H-1024_A-16 \
+WORD_DIM=1024 \
+LOWER_CASE=FALSE \
+BATCH_SIZE=4 \
+allennlp train \
+-s checkpoints/amr_bert \
+--include-package utils \
+--include-package modules \
+--file-friendly-logging \
+config/transition_bert_amr.jsonnet

config/transition_bert_amr.jsonnet

@@ -0,0 +1,81 @@
+{
+  "vocabulary": {
+    "non_padded_namespaces": []
+  },
+  "dataset_reader": {
+      "type": "amr_list-based_arc-eager",
+      "token_indexers": {
+        "tokens": {
+          "type": "bert-pretrained",
+          "pretrained_model": std.extVar('BERT_PATH'),
+          "do_lowercase": std.extVar('LOWER_CASE')=='TRUE'
+        }
+      }
+  },
+  "train_data_path": std.extVar('TRAIN_PATH'),
+  "validation_data_path": std.extVar('DEV_PATH'),
+  "model": {
+    "type": "transition_parser_amr",
+    "eval_on_training": false,
+    "text_field_embedder": {
+      "tokens": {
+        "type": "bert-pretrained",
+        "pretrained_model": std.extVar('BERT_PATH'),
+        "requires_grad": true,
+        "top_layer_only": false
+      },
+      "embedder_to_indexer_map": {
+        "tokens": ["tokens", "tokens-offsets", "tokens-type-ids"]
+      },
+      "allow_unmatched_keys": true
+    },
+    "word_dim": std.parseInt(std.extVar('WORD_DIM')),
+    "hidden_dim": 256,
+    "action_dim": 128,
+    "entity_dim": 64,
+    "rel_dim": 64,
+    "num_layers": 2,
+    "recurrent_dropout_probability": 0.2,
+    "layer_dropout_probability": 0.2,
+    "same_dropout_mask_per_instance": true,
+    "input_dropout": 0.2,
+    "initializer": [
+      ["p_.*weight", {"type": "xavier_uniform"}],
+      ["p_.*bias", {"type": "zero"}],
+      ["p(root|empty)_.*_emb", {"type": "normal"}],
+    ]
+  },
+  "iterator": {
+    "type": "bucket",
+    "sorting_keys": [["tokens", "num_tokens"]],
+    "batch_size": std.parseInt(std.extVar('BATCH_SIZE'))
+  },
+  "trainer": {
+    "num_epochs": 50,
+    "grad_norm": 5.0,
+    "grad_clipping": 5.0,
+    "patience": 50,
+    "cuda_device": 0,
+    "validation_metric": "+all-f",
+    "optimizer": {
+      "type": "adam",
+      "parameter_groups": [
+        [[".*bert.*"], {"lr": 5e-5}],
+        [["^((?!bert).)*$"], {}]
+      ],
+      "betas": [0.9, 0.999],
+      "lr": 1e-3
+    },
+    "learning_rate_scheduler": {
+      "type": "slanted_triangular",
+      "num_epochs": 50,
+      "num_steps_per_epoch": 1000,
+      "cut_frac": 0.1,
+      "ratio": 32,
+      "gradual_unfreezing": true,
+      "discriminative_fine_tuning": true,
+      "decay_factor": 1.0,
+    },
+    "num_serialized_models_to_keep": 50
+  }
+}