Relation Extraction

You can test the model at SGNLP-Demo.
If you want to find out more information, please contact us at [email protected].

Model Details
How to Get Started With the Model
Training
Model Parameters
Other Information

Model Details

Model Name: LSR

Description: This is a neural network that induces a latent document-level graph and uses a refinement strategy that allows the model to incrementally aggregate relevant information for multi-hop reasoning. This particular model corresponds to the GloVe+LSR model described in the paper.
Paper: Reasoning with Latent Structure Refinement for Document-Level Relation Extraction. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, July 2020 (pp. 1546-1557).
Author(s): Nan, G., Guo, Z., Sekulić, I., & Lu, W. (2020).
URL: https://aclanthology.org/2020.acl-main.141/

How to Get Started With the Model

Install Python package

SGnlp is an initiative by AI Singapore's NLP Hub. They aim to bridge the gap between research and industry, promote translational research, and encourage adoption of NLP techniques in the industry.

Various NLP models, other than relation extraction are available in the python package. You can try them out at SGNLP-Demo | SGNLP-Github.

pip install sgnlp

Examples

For more full code (such as Relation-Extraction), please refer to this github.
Alternatively, you can also try out the SGNLP-Demo | SGNLP-Docs for Relation extraction using LSR model.

Example of Relation Extraction (using LSR model):

from sgnlp.models.lsr import LsrModel, LsrConfig, LsrPreprocessor, LsrPostprocessor
from transformers import cached_path

# Download files from azure blob storage
rel2id_path = cached_path('https://storage.googleapis.com/sgnlp-models/models/lsr/rel2id.json')
word2id_path = cached_path('https://storage.googleapis.com/sgnlp-models/models/lsr/word2id.json')
ner2id_path = cached_path('https://storage.googleapis.com/sgnlp-models/models/lsr/ner2id.json')
rel_info_path = cached_path('https://storage.googleapis.com/sgnlp-models/models/lsr/rel_info.json')

PRED_THRESHOLD = 0.3
preprocessor = LsrPreprocessor(rel2id_path=rel2id_path, word2id_path=word2id_path, ner2id_path=ner2id_path)
postprocessor = LsrPostprocessor.from_file_paths(rel2id_path=rel2id_path, rel_info_path=rel_info_path,
                                                 pred_threshold=PRED_THRESHOLD)

# Load model
config = LsrConfig.from_pretrained('https://storage.googleapis.com/sgnlp-models/models/lsr/v2/config.json')
model = LsrModel.from_pretrained('https://storage.googleapis.com/sgnlp-models/models/lsr/v2/pytorch_model.bin', config=config)
model.eval()

# DocRED-like instance
instance = {
    "vertexSet": [[{"name": "Lark Force", "pos": [0, 2], "sent_id": 0, "type": "ORG"},
                   {"sent_id": 3, "type": "ORG", "pos": [2, 4], "name": "Lark Force"},
                   {"name": "Lark Force", "pos": [3, 5], "sent_id": 4, "type": "ORG"}],
                  [{"name": "Australian Army", "pos": [4, 6], "sent_id": 0, "type": "ORG"}],
                  [{"pos": [9, 11], "type": "TIME", "sent_id": 0, "name": "March 1941"}],
                  [{"name": "World War II", "pos": [12, 15], "sent_id": 0, "type": "MISC"}],
                  [{"name": "New Britain", "pos": [18, 20], "sent_id": 0, "type": "LOC"}],
                  [{"name": "New Ireland", "pos": [21, 23], "sent_id": 0, "type": "LOC"}],
                  [{"name": "John Scanlan", "pos": [6, 8], "sent_id": 1, "type": "PER"}],
                  [{"name": "Australia", "pos": [13, 14], "sent_id": 1, "type": "LOC"}],
                  [{"name": "Rabaul", "pos": [17, 18], "sent_id": 1, "type": "LOC"},
                   {"name": "Rabaul", "pos": [12, 13], "sent_id": 3, "type": "LOC"}],
                  [{"name": "Kavieng", "pos": [19, 20], "sent_id": 1, "type": "LOC"},
                   {"name": "Kavieng", "pos": [14, 15], "sent_id": 3, "type": "LOC"}],
                  [{"pos": [22, 24], "type": "MISC", "sent_id": 1, "name": "SS Katoomba"}],
                  [{"pos": [25, 27], "type": "MISC", "sent_id": 1, "name": "MV Neptuna"}],
                  [{"name": "HMAT Zealandia", "pos": [28, 30], "sent_id": 1, "type": "MISC"}],
                  [{"name": "Imperial Japanese Army", "pos": [8, 11], "sent_id": 3, "type": "ORG"}],
                  [{"pos": [18, 20], "type": "TIME", "sent_id": 3, "name": "January 1942"}],
                  [{"name": "Japan", "pos": [8, 9], "sent_id": 4, "type": "LOC"}],
                  [{"pos": [12, 13], "type": "MISC", "sent_id": 4, "name": "NCOs"}],
                  [{"name": "USS Sturgeon", "pos": [20, 22], "sent_id": 4, "type": "MISC"}],
                  [{"sent_id": 4, "type": "MISC", "pos": [27, 29], "name": "Montevideo Maru"}],
                  [{"name": "Japanese", "pos": [5, 6], "sent_id": 5, "type": "LOC"}],
                  [{"pos": [15, 16], "type": "NUM", "sent_id": 5, "name": "1,050"}],
                  [{"pos": [17, 18], "type": "NUM", "sent_id": 5, "name": "1,053"}]],
    "labels": [
        {"r": "P607", "h": 1, "t": 3, "evidence": [0]},
        {"r": "P17", "h": 1, "t": 7, "evidence": [0, 1]},
        {"r": "P241", "h": 6, "t": 1, "evidence": [0, 1]},
        {"r": "P607", "h": 6, "t": 3, "evidence": [0, 1]},
        {"r": "P27", "h": 6, "t": 7, "evidence": [0, 1]},
        {"r": "P1344", "h": 7, "t": 3, "evidence": [0, 1]},
        {"r": "P607", "h": 13, "t": 3, "evidence": [0, 3]},
        {"r": "P17", "h": 13, "t": 15, "evidence": [3, 4, 5]},
        {"r": "P17", "h": 13, "t": 19, "evidence": [3, 4, 5]},
        {"r": "P1344", "h": 15, "t": 3, "evidence": [0, 3, 4, 5]},
        {"r": "P172", "h": 15, "t": 19, "evidence": [4, 5]},
        {"r": "P607", "h": 17, "t": 3, "evidence": [0, 4]},
        {"r": "P17", "h": 11, "t": 7, "evidence": [1]},
        {"r": "P17", "h": 12, "t": 7, "evidence": [0, 1]},
        {"r": "P137", "h": 0, "t": 1, "evidence": [0, 1]},
        {"r": "P571", "h": 0, "t": 2, "evidence": [0]},
        {"r": "P607", "h": 0, "t": 3, "evidence": [0]},
        {"r": "P17", "h": 0, "t": 7, "evidence": [0, 1]}],
    "title": "Lark Force",
    "sents": [
        ["Lark", "Force", "was", "an", "Australian", "Army", "formation", "established", "in", "March", "1941",
         "during", "World", "War", "II", "for", "service", "in", "New", "Britain", "and", "New", "Ireland", "."],
        ["Under", "the", "command", "of", "Lieutenant", "Colonel", "John", "Scanlan", ",", "it", "was", "raised", "in",
         "Australia", "and", "deployed", "to", "Rabaul", "and", "Kavieng", ",", "aboard", "SS", "Katoomba", ",", "MV",
         "Neptuna", "and", "HMAT", "Zealandia", ",", "to", "defend", "their", "strategically", "important", "harbours",
         "and", "airfields", "."],
        ["The", "objective", "of", "the", "force", ",", "was", "to", "maintain", "a", "forward", "air", "observation",
         "line", "as", "long", "as", "possible", "and", "to", "make", "the", "enemy", "fight", "for", "this", "line",
         "rather", "than", "abandon", "it", "at", "the", "first", "threat", "as", "the", "force", "was", "considered",
         "too", "small", "to", "withstand", "any", "invasion", "."],
        ["Most", "of", "Lark", "Force", "was", "captured", "by", "the", "Imperial", "Japanese", "Army", "after",
         "Rabaul", "and", "Kavieng", "were", "captured", "in", "January", "1942", "."],
        ["The", "officers", "of", "Lark", "Force", "were", "transported", "to", "Japan", ",", "however", "the", "NCOs",
         "and", "men", "were", "unfortunately", "torpedoed", "by", "the", "USS", "Sturgeon", "while", "being",
         "transported", "aboard", "the", "Montevideo", "Maru", "."],
        ["Only", "a", "handful", "of", "the", "Japanese", "crew", "were", "rescued", ",", "with", "none", "of", "the",
         "between", "1,050", "and", "1,053", "prisoners", "aboard", "surviving", "as", "they", "were", "still",
         "locked", "below", "deck", "."]
    ]
}

tensor_doc = preprocessor([instance])
output = model(**tensor_doc)

result = postprocessor(output.prediction, [instance])

Training

The training datasets can be retrieved from Permuted dataset derived from Linguistic Data Consortium's (LDC) Wall Street Journal (WSJ) dataset. Please contact the authors to get the dataset if you have a valid LDC license.

Training Results

Training Time: ~17 hours for 100 epochs on a single V100 GPU.
Datasets: Retrieved from DocRED
Training Config: Not available.

Model Parameters

Model Weights: link
Model Config: link
Model Inputs: Coreference clusters of entities, relations between clusters of entities, and text.
Model Outputs: Scores of all possible relation labels between all possible pairs of entity clusters.
Model Size: ~85MB
Model Inference Info: Not available.
Usage Scenarios: Knowledge graph building.

Other Information

Original Code: link
Additional Information: CAVEATS: The model trained in this paper alone is not sufficient to do extract relations from a document. It requires other models to perform entity recognition and coreference between the entities. For this demo, two other pretrained models from AllenNLP is used: Fine Grained Name Entity Recognition and Coreference SpanBERT.

aisingapore
/

LSR