💬SPEECH: Structured Prediction with Energy-Based Event-Centric Hyperspheres

🍎 The project is an official implementation for SPEECH model and a repository for OntoEvent-Doc dataset, which has firstly been proposed in the paper 💬SPEECH: Structured Prediction with Energy-Based Event-Centric Hyperspheres accepted by ACL 2023 main conference.

🖥️ We also release the poster and slides for better understanding of this paper.

🤗 The implementations are based on Huggingface's Transformers and also referred to OntoED & DeepKE.

🤗 The baseline implementations are reproduced with codes referred to MAVEN's baselines or with official implementation.

SPEECH is proposed to address event-centric structured prediction with energy-based hyperspheres.
SPEECH models complex dependency among event structured components with energy-based modeling, and represents event classes with simple but effective hyperspheres.

The structure of data and code is as follows:

SPEECH
├── README.md
├── ACL2023@Poster_Speech.pdf
├── ACL2023@Slides_Speech.pdf
├── requirements.txt    # for package requirements
├── data_utils.py       # for data processing
├── speech.py		        # main model (bert serves as the backbone)
├── speech_distilbert.py	# main model (distilbert serves as the backbone)
├── speech_roberta.py           # toy model (roberta serves as the backbone, not adopted in the paper and just for reference)
├── run_speech.py	# for model running
├── run_speech.sh	# bash file for model running 
└── Datasets		    # data
    ├── MAVEN_ERE   
    │   ├── train.jsonl     # for training
    │   ├── test.jsonl      # for testing
    │   └── valid.jsonl     # for validation
    ├── OntoEvent-Doc
    │   ├── event_dict_label_data.json          # containing all event type labels 
    │   ├── event_dict_on_doc_train.json	# for training
    │   ├── event_dict_on_doc_test.json		# for testing
    │   └── event_dict_on_doc_valid.json	# for validation
    └── README.md 

• python==3.9.12

• torch==1.13.0

• transformers==4.25.1

• scikit-learn==1.2.2

• torchmetrics==0.9.3

• sentencepiece==0.1.97

1. Project Preparation:

Download this project and unzip the dataset. You can directly download the archive, or run git clone https://github.com/zjunlp/SPEECH.git in your teminal.

cd [LOCAL_PROJECT_PATH]

git clone [email protected]:zjunlp/SPEECH.git 

2. Data Preparation:

Unzip MAVEN_ERE and OntoEvent-Doc datasets stored at ./Datasets.

cd Datasets/
unzip MAVEN_ERE
unzip OntoEvent-Doc
cd .. 

3. Running Preparation:

Install all required packages.
Adjust the parameters in run_speech.sh bash file.

pip install -r requirements.txt
vim run_speech.sh
# input the parameters, save and quit

Hint:

• Please refer to main() function in run_speech.py file for detail meanings of each parameters.
• Pay attention to --ere_task_type parameter candidates:
  • "doc_all" is for "All Joint" experiments in the paper
  • "doc_joint" is for each ERE subtask "+joint" experiments in the paper
  • "doc_temporal"/"doc_causal/"doc_sub" is for each ERE subtask experiments only
• Note that the loss ratio λ1, λ2, λ3, for trigger classification, event classification and event-relation extraction depends on different tasks, please ensure a correct setting of these ratios, referring to line 56-61 in speech.py and speech_distilbert.py file for details. We also present the loss ratio setting in Appendix B in our paper.

4. Running Model:

Run ./run_speech.sh for training, validation, and testing.

./run_speech.sh

# Or you can run run_speech.py with manual parameter input in the terminal.

python run_speech.py --para... 

Hint:

• A folder of model checkpoints will be saved at the path you input (--output_dir) in the bash file run_speech.sh or the command line in the terminal.
• We also release the checkpoints for direct testing (Dismiss --do_train in the parameter input)

We briefly introduce the datasets in Section 4.1 and Appendix A in our paper.

MAVEN_ERE is proposed in a paper and released in GitHub.

OntoEvent-Doc, formatted in document level, is derived from OntoEvent which is formatted in sentence level.

The statistics of MAVEN-ERE and OntoEvent-Doc are shown below, and the detailed data schema can be referred to [./Datasets/README.md].

The data schema of MAVEN-ERE can be referred to their GitHub. Experiments on MAVEN-ERE in our paper involve:

• 6 temporal relations: BEFORE, OVERLAP, CONTAINS, SIMULTANEOUS, BEGINS-ON, ENDS-ON
• 2 causal relations: CAUSE, PRECONDITION
• 1 subevent relation: subevent_relations

Experiments on OntoEvent-Doc in our paper involve:

• 3 temporal relations: BEFORE, AFTER, EQUAL
• 2 causal relations: CAUSE, CAUSEDBY

We also add a NA relation to signify no relation between the event mention pair for the two datasets.

🍒 The OntoEvent-Doc dataset is stored in json format. Each document (specialized with a doc_id, e.g., 95dd35ce7dd6d377c963447eef47c66c) in OntoEvent-Doc datasets contains a list of "events" and a dictionary of "relations", where the data format is as below:

[a doc_id]:
{
    "events": [
    {
        'doc_id': '...', 
        'doc_title': 'XXX', 
        'sent_id': , 
        'event_mention': '......', 
        'event_mention_tokens': ['.', '.', '.', '.', '.', '.'], 
        'trigger': '...', 
        'trigger_pos': [, ], 
        'event_type': ''
    },
    {
        'doc_id': '...', 
        'doc_title': 'XXX', 
        'sent_id': , 
        'event_mention': '......', 
        'event_mention_tokens': ['.', '.', '.', '.', '.', '.'], 
        'trigger': '...', 
        'trigger_pos': [, ], 
        'event_type': ''
    },
    ... 
    ],
    "relations": { // each event-relation contains a list of 'sent_id' pairs.  
        "COSUPER": [[,], [,], [,]], 
        "SUBSUPER": [], 
        "SUPERSUB": [], 
        "CAUSE": [[,], [,]], 
        "BEFORE": [[,], [,]], 
        "AFTER": [[,], [,]], 
        "CAUSEDBY": [[,], [,]], 
        "EQUAL": [[,], [,]]
    }
} 

📋 Thank you very much for your interest in our work. If you use or extend our work, please cite the following paper:

@inproceedings{ACL2023_SPEECH,
    author    = {Shumin Deng and
                 Shengyu Mao and
                 Ningyu Zhang and
                 Bryan Hooi},
  title       = {SPEECH: Structured Prediction with Energy-Based Event-Centric Hyperspheres},
  booktitle   = {{ACL} {(1)}},
  publisher   = {Association for Computational Linguistics},
  pages       = {351--363},
  year        = {2023},
  url         = {https://aclanthology.org/2023.acl-long.21/}
}