Zero-Shot Transfer Learning for Event Extraction

Setup Instruction

This repository is built on Theano (=0.8, verifed. Issues may be reported on recent versions), Python 2.7. The environment can be quickly set up by running

pip install -r requirements.txt

Run the Model

Step 1

Preprocess Data: remove XML tags from ACE/ERE articles, sentence merging, sentence segmantation and tokenization.

java -jar removeXmlTag.jar [source_path] [source_without_tag_path]

python rsd2ltf.py [source_without_tag_path] [ltf_path] --extension .txt --seg_option nltk+linebreak

java -jar sentenceExtractor.jar [ltf_path] [sent_path] [sent_id_path]

[source_path]: the path for input files (all .sgm files from ACE source corpus)

[source_without_tag_path]: output path after merging sentences and removing xml tags from ACE source articles

[ltf_path]: output path after sentence segmentation and tokenization

[sent_path] [sent_id_path]: output paths for sentences and sentence ids

Step 2

Apply AMR parser (https://github.com/c-amr/camr, or https://github.com/jflanigan/jamr) to parse each single doc, and get aligned parsing output [amr_parsing_path].

Step 3

Extract event mention structures from AMR parsing outputs

java -jar amrPostProcessing.jar [amr_parsing_path] [sent_id_path] [resource/amrRelationsAnnotated.txt] [resource/frameNetVN.txt] [trigger_structure_file]

[amr_parsing_path]: path of AMR parsing files generated from Step2

[sent_id_path]: path of sentence ids generated from Step1

[resource/amrRelationsAnnotated.txt]: path for resource/amrRelationsAnnotated.txt file

[resource/frameNetVN.txt]: path for resource/frameNetVN.txt file

[trigger_structure_file]: output file containing all event structures extracted from all the input files

Step 4

Apply zero shot trigger extraction model

python zero_shot_final_test.py --test [trigger_structure_file] --test_result [trigger_prediction_file] --model_path [model_file] --embedding_path [embedding_file] --ontology_path [ontology_file] --norm_ontology_path [normalized_ontology_file] --seen_types [seen_type_file] --relation_path [file_of_amr_relations]

To retrain the model, you can update the arguments in zero_shot_final.py file. The ACE training data can be generated from the data/sample/aceEventStructure.train.format.pos_neg.txt

[trigger_structure_file]: input file containing all event structures generated from Step3

[trigger_prediction_file]: output file containing all event structures and predicted types

[embedding_file]: e.g., data/embedding/wsd.model.ace.filter.txt

[ontology_file]: file with the whole target event ontology, e.g., if you use Framenet+ACE as the target ontology, it should be data/frame-ontology/event.ontology.new.txt

[normalized_ontology_file]: e.g., data/frame-ontology/event.ontology.normalize.new.txt

[seen_type_file]: file with all seen/training types, e.g., data/flags/train.10

[file_of_amr_relations]: file with all meaningful AMR relations, e.g., data/resource/amrRelations.txt

Step 5

Extract candidate arguments

java -jar prepareArgPrediction.jar [amr_parsing_path] [trigger_structure_file] [trigger_prediction_file] [resource/frame-ontology/event.ontology.new.txt] [arg_structure_file]

[amr_parsing_path]: path of AMR parsing files generated from Step2

[trigger_structure_file]: path of event structure file generated from Step3

[trigger_prediction_file]: path of trigger prediction file generated from Step4

[resource/frame-ontology/event.ontology.new.txt]: path to the file resource/frame-ontology/event.ontology.new.txt

[arg_structure_file]: output file containing all argument structures

Step 6

Apply zero shot argument extraction model

python zero_shot_arg_final_test.py --test [arg_structure_file] --test_result [arg_prediction_file] --model_path [model_file] --embedding_path [embedding_file] --ontology_path [ontology_file] --norm_ontology_path [normalized_ontology_file] --arg_path_file [argument_paths_file] --arg_path_file_universal [universal_argument_paths_file] --trigger_role_matrix [trigger_role_masks] --seen_argss [seen_argument_roles_file] --relation_path [file_of_amr_relations]

[arg_structure_file]: argument structure file generated from Step5

[arg_prediction_file]: output file containing all argument structures and argument role predictions

[argument_paths_file]: list of all event types and argument roles, e.g., data/frame-ontology/event.ontology.args.merge.all.txt

[universal_argument_paths_file]: list of all argument roles shared by all event types, e.g., data/frame-ontology/event.ontology.args.universal.txt

[trigger_role_masks]: matrix to show each event types (index) with all predefined argument roles (index), e.g., data/frame-ontology/event.ontology.trigger.arg.matrix.new.txt

[seen_argument_roles_file]: all seens/training argument roles, e.g., data/flags/train.arg.10

Citation

[1] Lifu Huang, Heng Ji, Kyunghyun Cho, Clare R Voss. Zero-shot transfer learning for event extraction, Proc. ACL, 2018