Fast and Accurate Entity Linking and Relation Extraction on an Academic Budget
Project description
Retrieve, Read and LinK: Fast and Accurate Entity Linking and Relation Extraction on an Academic Budget
A blazing fast and lightweight Information Extraction model for Entity Linking and Relation Extraction.
๐ ๏ธ Installation
Installation from PyPI
pip install relik
Other installation options
Install with optional dependencies
Install with all the optional dependencies.
pip install relik[all]
Install with optional dependencies for training and evaluation.
pip install relik[train]
Install with optional dependencies for FAISS
FAISS PyPI package is only available for CPU. For GPU, install it from source or use the conda package.
For CPU:
pip install relik[faiss]
For GPU:
conda create -n relik python=3.10
conda activate relik
# install pytorch
conda install -y pytorch=2.1.0 pytorch-cuda=12.1 -c pytorch -c nvidia
# GPU
conda install -y -c pytorch -c nvidia faiss-gpu=1.8.0
# or GPU with NVIDIA RAFT
conda install -y -c pytorch -c nvidia -c rapidsai -c conda-forge faiss-gpu-raft=1.8.0
pip install relik
Install with optional dependencies for serving the models with FastAPI and Ray.
pip install relik[serve]
Installation from source
git clone https://github.com/SapienzaNLP/relik.git
cd relik
pip install -e .[all]
๐ค Models
-
ReLiK Large for Relation Extraction (๐ Large RE v2, Colab โ ):
relik-ie/relik-relation-extraction-large -
ReLiK Large for Closed Information Extraction (๐ large EL + RE, Colab โ ):
relik-ie/relik-relation-extraction-large -
ReLiK Extra Large for Closed Information Extraction (๐ Our thicc boi for EL + RE):
relik-ie/relik-cie-xl -
ReLiK Small for Entity Linking (๐๐คโก Tiny and Fast EL, Colab โ ):
sapienzanlp/relik-entity-linking-small -
ReLiK Small for Entity Linking (โก Small and Fast EL):
sapienzanlp/relik-entity-linking-small -
ReLiK Small for Closed Information Extraction (๐ฅ EL + RE):
relik-ie/relik-cie-small -
ReLiK Large for Entity Linking (๐ฅ EL for the wild):
relik-ie/relik-entity-linking-large-robust -
ReLiK Small for Entity Linking (๐ฅ RE + NER):
relik-ie/relik-relation-extraction-small-wikipedia-ner
Models from the paper:
- ReLiK Large for Entity Linking (๐ Paper version):
sapienzanlp/relik-entity-linking-large - ReLik Base for Entity Linking (๐ Paper version):
sapienzanlp/relik-entity-linking-base - ReLiK Large for Relation Extraction (๐ Paper version):
sapienzanlp/relik-relation-extraction-nyt-large
A full list of models can be found on ๐ค Hugging Face.
Other models sizes will be available in the future ๐.
๐ Quick Start
ReLiK is a lightweight and fast model for Entity Linking and Relation Extraction.
It is composed of two main components: a retriever and a reader.
The retriever is responsible for retrieving relevant documents from a large collection,
while the reader is responsible for extracting entities and relations from the retrieved documents.
ReLiK can be used with the from_pretrained method to load a pre-trained pipeline.
Here is an example of how to use ReLiK for Entity Linking:
from relik import Relik
from relik.inference.data.objects import RelikOutput
relik = Relik.from_pretrained("sapienzanlp/relik-entity-linking-large")
relik_out: RelikOutput = relik("Michael Jordan was one of the best players in the NBA.")
Output:
RelikOutput(
text="Michael Jordan was one of the best players in the NBA.",
tokens=['Michael', 'Jordan', 'was', 'one', 'of', 'the', 'best', 'players', 'in', 'the', 'NBA', '.'],
id=0,
spans=[
Span(start=0, end=14, label="Michael Jordan", text="Michael Jordan"),
Span(start=50, end=53, label="National Basketball Association", text="NBA"),
],
triples=[],
candidates=Candidates(
span=[
[
[
{"text": "Michael Jordan", "id": 4484083},
{"text": "National Basketball Association", "id": 5209815},
{"text": "Walter Jordan", "id": 2340190},
{"text": "Jordan", "id": 3486773},
{"text": "50 Greatest Players in NBA History", "id": 1742909},
...
]
]
]
),
)
and for Relation Extraction:
from relik import Relik
from relik.inference.data.objects import RelikOutput
relik = Relik.from_pretrained("sapienzanlp/relik-relation-extraction-nyt-large")
relik_out: RelikOutput = relik("Michael Jordan was one of the best players in the NBA.")
Output:
RelikOutput(
text='Michael Jordan was one of the best players in the NBA.',
tokens=Michael Jordan was one of the best players in the NBA.,
id=0,
spans=[
Span(start=0, end=14, label='--NME--', text='Michael Jordan'),
Span(start=50, end=53, label='--NME--', text='NBA')
],
triplets=[
Triplets(
subject=Span(start=0, end=14, label='--NME--', text='Michael Jordan'),
label='company',
object=Span(start=50, end=53, label='--NME--', text='NBA'),
confidence=1.0
)
],
candidates=Candidates(
span=[],
triplet=[
[
[
{"text": "company", "id": 4, "metadata": {"definition": "company of this person"}},
{"text": "nationality", "id": 10, "metadata": {"definition": "nationality of this person or entity"}},
{"text": "child", "id": 17, "metadata": {"definition": "child of this person"}},
{"text": "founded by", "id": 0, "metadata": {"definition": "founder or co-founder of this organization, religion or place"}},
{"text": "residence", "id": 18, "metadata": {"definition": "place where this person has lived"}},
...
]
]
]
),
)
Usage
Retrievers and Readers can be used separately. In the case of retriever-only ReLiK, the output will contain the candidates for the input text.
Retriever-only example:
from relik import Relik
from relik.inference.data.objects import RelikOutput
# If you want to use only the retriever
retriever = Relik.from_pretrained("sapienzanlp/relik-entity-linking-large", reader=None)
relik_out: RelikOutput = retriever("Michael Jordan was one of the best players in the NBA.")
Output:
RelikOutput(
text="Michael Jordan was one of the best players in the NBA.",
tokens=['Michael', 'Jordan', 'was', 'one', 'of', 'the', 'best', 'players', 'in', 'the', 'NBA', '.'],
id=0,
spans=[],
triples=[],
candidates=Candidates(
span=[
[
{"text": "Michael Jordan", "id": 4484083},
{"text": "National Basketball Association", "id": 5209815},
{"text": "Walter Jordan", "id": 2340190},
{"text": "Jordan", "id": 3486773},
{"text": "50 Greatest Players in NBA History", "id": 1742909},
...
]
],
triplet=[],
),
)
Reader-only example:
from relik import Relik
from relik.inference.data.objects import RelikOutput
# If you want to use only the reader
reader = Relik.from_pretrained("sapienzanlp/relik-entity-linking-large", retriever=None)
candidates = [
"Michael Jordan",
"National Basketball Association",
"Walter Jordan",
"Jordan",
"50 Greatest Players in NBA History",
]
text = "Michael Jordan was one of the best players in the NBA."
relik_out: RelikOutput = reader(text, candidates=candidates)
Output:
RelikOutput(
text="Michael Jordan was one of the best players in the NBA.",
tokens=['Michael', 'Jordan', 'was', 'one', 'of', 'the', 'best', 'players', 'in', 'the', 'NBA', '.'],
id=0,
spans=[
Span(start=0, end=14, label="Michael Jordan", text="Michael Jordan"),
Span(start=50, end=53, label="National Basketball Association", text="NBA"),
],
triples=[],
candidates=Candidates(
span=[
[
[
{
"text": "Michael Jordan",
"id": -731245042436891448,
},
{
"text": "National Basketball Association",
"id": 8135443493867772328,
},
{
"text": "Walter Jordan",
"id": -5873847607270755146,
"metadata": {},
},
{"text": "Jordan", "id": 6387058293887192208, "metadata": {}},
{
"text": "50 Greatest Players in NBA History",
"id": 2173802663468652889,
},
]
]
],
),
)
CLI
ReLiK provides a CLI to serve a FastAPI server for the model or to perform inference on a dataset.
relik serve
relik serve --help
Usage: relik serve [OPTIONS] RELIK_PRETRAINED [DEVICE] [RETRIEVER_DEVICE]
[DOCUMENT_INDEX_DEVICE] [READER_DEVICE] [PRECISION]
[RETRIEVER_PRECISION] [DOCUMENT_INDEX_PRECISION]
[READER_PRECISION] [ANNOTATION_TYPE]
โญโ Arguments โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ * relik_pretrained TEXT [default: None] [required] โ
โ device [DEVICE] The device to use for relik (e.g., โ
โ 'cuda', 'cpu'). โ
โ [default: None] โ
โ retriever_device [RETRIEVER_DEVICE] The device to use for the retriever โ
โ (e.g., 'cuda', 'cpu'). โ
โ [default: None] โ
โ document_index_device [DOCUMENT_INDEX_DEVICE] The device to use for the index โ
โ (e.g., 'cuda', 'cpu'). โ
โ [default: None] โ
โ reader_device [READER_DEVICE] The device to use for the reader โ
โ (e.g., 'cuda', 'cpu'). โ
โ [default: None] โ
โ precision [PRECISION] The precision to use for relik โ
โ (e.g., '32', '16'). โ
โ [default: 32] โ
โ retriever_precision [RETRIEVER_PRECISION] The precision to use for the โ
โ retriever (e.g., '32', '16'). โ
โ [default: None] โ
โ document_index_precision [DOCUMENT_INDEX_PRECISION] The precision to use for the index โ
โ (e.g., '32', '16'). โ
โ [default: None] โ
โ reader_precision [READER_PRECISION] The precision to use for the reader โ
โ (e.g., '32', '16'). โ
โ [default: None] โ
โ annotation_type [ANNOTATION_TYPE] The type of annotation to use (e.g., โ
โ 'CHAR', 'WORD'). โ
โ [default: char] โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
โญโ Options โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ --host TEXT [default: 0.0.0.0] โ
โ --port INTEGER [default: 8000] โ
โ --frontend --no-frontend [default: no-frontend] โ
โ --help Show this message and exit. โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
For example:
relik serve sapienzanlp/relik-entity-linking-large
relik inference
relik inference --help
Usage: relik inference [OPTIONS] MODEL_NAME_OR_PATH INPUT_PATH OUTPUT_PATH
โญโ Arguments โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ * model_name_or_path TEXT [default: None] [required] โ
โ * input_path TEXT [default: None] [required] โ
โ * output_path TEXT [default: None] [required] โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
โญโ Options โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ --batch-size INTEGER [default: 8] โ
โ --num-workers INTEGER [default: 4] โ
โ --device TEXT [default: cuda] โ
โ --precision TEXT [default: fp16] โ
โ --top-k INTEGER [default: 100] โ
โ --window-size INTEGER [default: None] โ
โ --window-stride INTEGER [default: None] โ
โ --annotation-type TEXT [default: char] โ
โ --progress-bar --no-progress-bar [default: progress-bar] โ
โ --model-kwargs TEXT [default: None] โ
โ --inference-kwargs TEXT [default: None] โ
โ --help Show this message and exit. โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
For example:
relik inference sapienzanlp/relik-entity-linking-large data.txt output.jsonl
Docker Images
Docker images for ReLiK are available on Docker Hub. You can pull the latest image with:
docker pull sapienzanlp/relik:latest
and run the image with:
docker run -p 12345:8000 sapienzanlp/relik:latest -c relik-ie/relik-cie-small
The API will be available at http://localhost:12345. It exposes a single endpoint /relik with several parameters that can be passed to the model.
A quick documentation of the API can be found at http://localhost:12345/docs. Here is a simple example of how to query the API:
curl -X 'GET' \
'http://127.0.0.1:12345/api/relik?text=Michael%20Jordan%20was%20one%20of%20the%20best%20players%20in%20the%20NBA.&is_split_into_words=false&retriever_batch_size=32&reader_batch_size=32&return_windows=false&use_doc_topic=false&annotation_type=char&relation_threshold=0.5' \
-H 'accept: application/json'
Here the full list of parameters that can be passed to the docker image:
docker run sapienzanlp/relik:latest -h
Usage: relik [-h --help] [-c --config] [-p --precision] [-d --device] [--retriever] [--retriever-device]
[--retriever-precision] [--index-device] [--index-precision] [--reader] [--reader-device] [--reader-precision]
[--annotation-type] [--frontend] [--workers] -- start the FastAPI server for the RElik model
where:
-h --help Show this help text
-c --config Pretrained ReLiK config name (from HuggingFace) or path
-p --precision Precision, default '32'.
-d --device Device to use, default 'cpu'.
--retriever Override retriever model name.
--retriever-device Override retriever device.
--retriever-precision Override retriever precision.
--index-device Override index device.
--index-precision Override index precision.
--reader Override reader model name.
--reader-device Override reader device.
--reader-precision Override reader precision.
--annotation-type Annotation type ('char', 'word'), default 'char'.
--frontend Whether to start the frontend server.
--workers Number of workers to use.
๐ Before You Start
In the following sections, we provide a step-by-step guide on how to prepare the data, train the retriever and reader, and evaluate the model.
Entity Linking
All your data should have the following structure:
{
"doc_id": int, # Unique identifier for the document
"doc_text": txt, # Text of the document
"doc_span_annotations": # Char level annotations
[
[start, end, label],
[start, end, label],
...
]
}
We used BLINK (Wu et al., 2019) and AIDA (Hoffart et al, 2011) datasets for training and evaluation. More specifically, we used the BLINK dataset for pre-training the retriever and the AIDA dataset for fine-tuning the retriever and training the reader.
The BLINK dataset can be downloaded from the GENRE repo using this
script.
We used blink-train-kilt.jsonl and blink-dev-kilt.jsonl as training and validation datasets.
Assuming we have downloaded the two files in the data/blink folder, we converted the BLINK dataset to the ReLiK format using the following script:
# Train
python scripts/data/blink/preprocess_genre_blink.py \
data/blink/blink-train-kilt.jsonl \
data/blink/processed/blink-train-kilt-relik.jsonl
# Dev
python scripts/data/blink/preprocess_genre_blink.py \
data/blink/blink-dev-kilt.jsonl \
data/blink/processed/blink-dev-kilt-relik.jsonl
The AIDA dataset is not publicly available, but we provide the file we used without text field. You can find the file in ReLiK format in data/aida/processed folder.
The Wikipedia index we used can be downloaded from here.
Relation Extraction
All your data should have the following structure:
{
"doc_id": int, # Unique identifier for the document
"doc_words: list[txt] # Tokenized text of the document
"doc_span_annotations": # Token level annotations of mentions (label is optional)
[
[start, end, label],
[start, end, label],
...
],
"doc_triplet_annotations": # Triplet annotations
[
{
"subject": [start, end, label], # label is optional
"relation": name, # type is optional
"object": [start, end, label], # label is optional
},
{
"subject": [start, end, label], # label is optional
"relation": name, # type is optional
"object": [start, end, label], # label is optional
},
]
}
For Relation Extraction, we provide an example of how to preprocess the NYT dataset from raw_nyt taken from CopyRE. Download the dataset to data/raw_nyt and then run:
python scripts/data/nyt/preprocess_nyt.py data/raw_nyt data/nyt/processed/
Please be aware that for fair comparison we reproduced the preprocessing from previous work, which leads to duplicate triplets due to the incorrect handling of repeated surface forms for entity spans. If you want to correctly parse the original data to ReLiK format, you can set the flag --legacy-format False. Just be aware that the provided RE NYT models were trained on the legacy format.
๐ฆฎ Retriever
We perform a two-step training process for the retriever. First, we "pre-train" the retriever using BLINK (Wu et al., 2019) dataset, and then we "fine-tune" it using AIDA (Hoffart et al, 2011).
Data Preparation
The retriever requires a dataset in a format similar to DPR: a jsonl file where each line is a dictionary with the following keys:
{
"question": "....",
"positive_ctxs": [{
"title": "...",
"text": "...."
}],
"negative_ctxs": [{
"title": "...",
"text": "...."
}],
"hard_negative_ctxs": [{
"title": "...",
"text": "...."
}]
}
The retriever also needs an index to search for the documents. The documents to index can be either a JSONL file or a TSV file similar to DPR:
jsonl: each line is a JSON object with the following keys:id,text,metadatatsv: each line is a tab-separated string with theidandtextcolumns, followed by any other column that will be stored in themetadatafield
jsonl example:
{
"id": "...",
"text": "...",
"metadata": ["{...}"]
},
...
tsv example:
id \t text \t any other column
...
Entity Linking
BLINK
Once you have the BLINK dataset in the ReLiK format, you can create the windows with the following script:
# train
relik data create-windows \
data/blink/processed/blink-train-kilt-relik.jsonl \
data/blink/processed/blink-train-kilt-relik-windowed.jsonl
# dev
relik data create-windows \
data/blink/processed/blink-dev-kilt-relik.jsonl \
data/blink/processed/blink-dev-kilt-relik-windowed.jsonl
and then convert it to the DPR format:
# train
relik data convert-to-dpr \
data/blink/processed/blink-train-kilt-relik-windowed.jsonl \
data/blink/processed/blink-train-kilt-relik-windowed-dpr.jsonl \
data/kb/wikipedia/documents.jsonl \
--title-map data/kb/wikipedia/title_map.json
# dev
relik data convert-to-dpr \
data/blink/processed/blink-dev-kilt-relik-windowed.jsonl \
data/blink/processed/blink-dev-kilt-relik-windowed-dpr.jsonl \
data/kb/wikipedia/documents.jsonl \
--title-map data/kb/wikipedia/title_map.json
AIDA
Since the AIDA dataset is not publicly available, we can provide the annotations for the AIDA dataset in the ReLiK format as an example.
Assuming you have the full AIDA dataset in the data/aida, you can convert it to the ReLiK format and then create the windows with the following script:
relik data create-windows \
data/aida/processed/aida-train-relik.jsonl \
data/aida/processed/aida-train-relik-windowed.jsonl
and then convert it to the DPR format:
relik data convert-to-dpr \
data/aida/processed/aida-train-relik-windowed.jsonl \
data/aida/processed/aida-train-relik-windowed-dpr.jsonl \
data/kb/wikipedia/documents.jsonl \
--title-map data/kb/wikipedia/title_map.json
Relation Extraction
NYT
relik data create-windows \
data/data/processed/nyt/train.jsonl \
data/data/processed/nyt/train-windowed.jsonl \
--is-split-into-words \
--window-size none
and then convert it to the DPR format:
relik data convert-to-dpr \
data/data/processed/nyt/train-windowed.jsonl \
data/data/processed/nyt/train-windowed-dpr.jsonl
Training the model
The relik retriever train command can be used to train the retriever. It requires the following arguments:
config_path: The path to the configuration file.overrides: A list of overrides to the configuration file, in the formatkey=value.
Examples of configuration files can be found in the relik/retriever/conf folder.
Entity Linking
The configuration files in relik/retriever/conf are pretrain_iterable_in_batch.yaml and finetune_iterable_in_batch.yaml, which we used to pre-train and fine-tune the retriever, respectively.
For instance, to train the retriever on the AIDA dataset, you can run the following command:
relik retriever train relik/retriever/conf/finetune_iterable_in_batch.yaml \
model.language_model=intfloat/e5-base-v2 \
data.train_dataset_path=data/aida/processed/aida-train-relik-windowed-dpr.jsonl \
data.val_dataset_path=data/aida/processed/aida-dev-relik-windowed-dpr.jsonl \
data.test_dataset_path=data/aida/processed/aida-test-relik-windowed-dpr.jsonl \
data.shared_params.documents_path=data/kb/wikipedia/documents.jsonl
Relation Extraction
The configuration file in relik/retriever/conf is finetune_nyt_iterable_in_batch.yaml, which we used to fine-tune the retriever for the NYT dataset. For cIE we repurpose the one pretrained from BLINK in the previous step.
For instance, to train the retriever on the NYT dataset, you can run the following command:
relik retriever train relik/retriever/conf/finetune_nyt_iterable_in_batch.yaml \
model.language_model=intfloat/e5-base-v2 \
data.train_dataset_path=data/nyt/processed/nyt-train-relik-windowed-dpr.jsonl \
data.val_dataset_path=data/nyt/processed/nyt-dev-relik-windowed-dpr.jsonl \
data.test_dataset_path=data/nyt/processed/nyt-test-relik-windowed-dpr.jsonl
Inference
By passing train.only_test=True to the relik retriever train command, you can skip the training and only evaluate the model.
It needs also the path to the PyTorch Lightning checkpoint and the dataset to evaluate on.
relik retriever train relik/retriever/conf/finetune_iterable_in_batch.yaml \
train.only_test=True \
test_dataset_path=data/aida/processed/aida-test-relik-windowed-dpr.jsonl
model.checkpoint_path=path/to/checkpoint
The retriever encoder can be saved from the checkpoint with the following command:
from relik.retriever.lightning_modules.pl_modules import GoldenRetrieverPLModule
checkpoint_path = "path/to/checkpoint"
retriever_folder = "path/to/retriever"
# If you want to push the model to the Hugging Face Hub set push_to_hub=True
push_to_hub = False
# If you want to push the model to the Hugging Face Hub set the repo_id
repo_id = "sapienzanlp/relik-retriever-e5-base-v2-aida-blink-encoder"
pl_module = GoldenRetrieverPLModule.load_from_checkpoint(checkpoint_path)
pl_module.model.save_pretrained(retriever_folder, push_to_hub=push_to_hub, repo_id=repo_id)
With push_to_hub=True the model will be pushed to the ๐ค Hugging Face Hub with repo_id as the repository id where the model will be pushed.
The retriever needs an index to search for the documents. The index can be created using relik retriever build-index command
relik retriever build-index --help
Usage: relik retriever build-index [OPTIONS] QUESTION_ENCODER_NAME_OR_PATH
DOCUMENT_PATH OUTPUT_FOLDER
โญโ Arguments โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ * question_encoder_name_or_path TEXT [default: None] [required] โ
โ * document_path TEXT [default: None] [required] โ
โ * output_folder TEXT [default: None] [required] โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
โญโ Options โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ --document-file-type TEXT [default: jsonl] โ
โ --passage-encoder-name-or-path TEXT [default: None] โ
โ --indexer-class TEXT [default: relik.retriever.indexers.inmemory.InMemoryDocumentIndex] โ
โ --batch-size INTEGER [default: 512] โ
โ --num-workers INTEGER [default: 4] โ
โ --passage-max-length INTEGER [default: 64] โ
โ --device TEXT [default: cuda] โ
โ --index-device TEXT [default: cpu] โ
โ --precision TEXT [default: fp32] โ
โ --push-to-hub --no-push-to-hub [default: no-push-to-hub] โ
โ --repo-id TEXT [default: None] โ
โ --help Show this message and exit. โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
With the encoder and the index, the retriever can be loaded from a repo id or a local path:
from relik.retriever import GoldenRetriever
encoder_name_or_path = "sapienzanlp/relik-retriever-e5-base-v2-aida-blink-encoder"
index_name_or_path = "sapienzanlp/relik-retriever-e5-base-v2-aida-blink-wikipedia-index"
retriever = GoldenRetriever(
question_encoder=encoder_name_or_path,
document_index=index_name_or_path,
device="cuda", # or "cpu"
precision="16", # or "32", "bf16"
index_device="cuda", # or "cpu"
index_precision="16", # or "32", "bf16"
)
and then it can be used to retrieve documents:
retriever.retrieve("Michael Jordan was one of the best players in the NBA.", top_k=100)
๐ค Reader
The reader is responsible for extracting entities and relations from documents from a set of candidates (e.g., possible entities or relations).
The reader can be trained for span extraction or triplet extraction.
The RelikReaderForSpanExtraction is used for span extraction, i.e. Entity Linking, while the RelikReaderForTripletExtraction is used for triplet extraction, i.e. Relation Extraction.
Data Preparation
The reader requires the windowized dataset we created in Section Before You Start augmented with the candidates from the retriever.
The candidates can be added to the dataset using the relik retriever add-candidates command.
relik retriever add-candidates --help
Usage: relik retriever add-candidates [OPTIONS] QUESTION_ENCODER_NAME_OR_PATH
DOCUMENT_NAME_OR_PATH INPUT_PATH
OUTPUT_PATH
โญโ Arguments โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ * question_encoder_name_or_path TEXT [default: None] [required] โ
โ * document_name_or_path TEXT [default: None] [required] โ
โ * input_path TEXT [default: None] [required] โ
โ * output_path TEXT [default: None] [required] โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
โญโ Options โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฎ
โ --passage-encoder-name-or-path TEXT [default: None] โ
โ --relations BOOLEAN [default: False] โ
โ --top-k INTEGER [default: 100] โ
โ --batch-size INTEGER [default: 128] โ
โ --num-workers INTEGER [default: 4] โ
โ --device TEXT [default: cuda] โ
โ --index-device TEXT [default: cpu] โ
โ --precision TEXT [default: fp32] โ
โ --use-doc-topics --no-use-doc-topics [default: no-use-doc-topics] โ
โ --help Show this message and exit. โ
โฐโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโฏ
Entity Linking
We need to add candidates to each window that will be used by the Reader, using our previously trained Retriever. Here is an example using our already trained retriever on Aida for the train split:
relik retriever add-candidates sapienzanlp/relik-retriever-e5-base-v2-aida-blink-encoder sapienzanlp/relik-retriever-e5-base-v2-aida-blink-wikipedia-index data/aida/processed/aida-train-relik-windowed.jsonl data/aida/processed/aida-train-relik-windowed-candidates.jsonl
Relation Extraction
The same thing happens for Relation Extraction. If you want to use our trained retriever:
relik retriever add-candidates sapienzanlp/relik-retriever-small-nyt-question-encoder sapienzanlp/relik-retriever-small-nyt-document-index data/nyt/processed/nyt-train-relik-windowed.jsonl data/nyt/processed/nyt-train-relik-windowed-candidates.jsonl
Training the model
Similar to the retriever, the relik reader train command can be used to train the retriever. It requires the following arguments:
config_path: The path to the configuration file.overrides: A list of overrides to the configuration file, in the formatkey=value.
Examples of configuration files can be found in the relik/reader/conf folder.
Entity Linking
The configuration files in relik/reader/conf are large.yaml and base.yaml, which we used to train the large and base reader, respectively.
For instance, to train the large reader on the AIDA dataset run:
relik reader train relik/reader/conf/large.yaml \
train_dataset_path=data/aida/processed/aida-train-relik-windowed-candidates.jsonl \
val_dataset_path=data/aida/processed/aida-dev-relik-windowed-candidates.jsonl \
test_dataset_path=data/aida/processed/aida-dev-relik-windowed-candidates.jsonl
Relation Extraction
The configuration files in relik/reader/conf are large_nyt.yaml, base_nyt.yaml, and small_nyt.yaml, which we used to train the large, base and small reader, respectively.
For instance, to train the large reader on the AIDA dataset run:
relik reader train relik/reader/conf/large_nyt.yaml \
train_dataset_path=data/nyt/processed/nyt-train-relik-windowed-candidates.jsonl \
val_dataset_path=data/nyt/processed/nyt-dev-relik-windowed-candidates.jsonl \
test_dataset_path=data/nyt/processed/nyt-test-relik-windowed-candidates.jsonl
Inference
The reader can be saved from the checkpoint with the following command:
from relik.reader.lightning_modules.relik_reader_pl_module import RelikReaderPLModule
checkpoint_path = "path/to/checkpoint"
reader_folder = "path/to/reader"
# If you want to push the model to the Hugging Face Hub set push_to_hub=True
push_to_hub = False
# If you want to push the model to the Hugging Face Hub set the repo_id
repo_id = "sapienzanlp/relik-reader-deberta-v3-large-aida"
pl_model = RelikReaderPLModule.load_from_checkpoint(
trainer.checkpoint_callback.best_model_path
)
pl_model.relik_reader_core_model.save_pretrained(experiment_path, push_to_hub=push_to_hub, repo_id=repo_id)
with push_to_hub=True the model will be pushed to the ๐ค Hugging Face Hub with repo_id as the repository id where the model will be uploaded.
The reader can be loaded from a repo id or a local path:
from relik.reader import RelikReaderForSpanExtraction, RelikReaderForTripletExtraction
# the reader for span extraction
reader_span = RelikReaderForSpanExtraction(
"sapienzanlp/relik-reader-deberta-v3-large-aida"
)
# the reader for triplet extraction
reader_tripltes = RelikReaderForTripletExtraction(
"sapienzanlp/relik-reader-deberta-v3-large-nyt"
)
and used to extract entities and relations:
# an example of candidates for the reader
candidates = ["Michael Jordan", "NBA", "Chicago Bulls", "Basketball", "United States"]
reader_span.read("Michael Jordan was one of the best players in the NBA.", candidates=candidates)
๐ Performance
Entity Linking
We evaluate the performance of ReLiK on Entity Linking using GERBIL. The following table shows the results (InKB Micro F1) of ReLiK Large and Base:
| Model | AIDA | MSNBC | Der | K50 | R128 | R500 | O15 | O16 | Tot | OOD | AIT (m:s) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| GENRE | 83.7 | 73.7 | 54.1 | 60.7 | 46.7 | 40.3 | 56.1 | 50.0 | 58.2 | 54.5 | 38:00 |
| EntQA | 85.8 | 72.1 | 52.9 | 64.5 | 54.1 | 41.9 | 61.1 | 51.3 | 60.5 | 56.4 | 20:00 |
| ReLiKsmall | 82.2 | 72.7 | 55.6 | 68.3 | 48.0 | 42.3 | 62.7 | 53.6 | 60.7 | 57.6 | 00:29 |
| ReLiKBase | 85.3 | 72.3 | 55.6 | 68.0 | 48.1 | 41.6 | 62.5 | 52.3 | 60.7 | 57.2 | 00:29 |
| ReLiKLarge | 86.4 | 75.0 | 56.3 | 72.8 | 51.7 | 43.0 | 65.1 | 57.2 | 63.4 | 60.2 | 01:46 |
Comparison systems' evaluation (InKB Micro F1) on the in-domain AIDA test set and out-of-domain MSNBC (MSN), Derczynski (Der), KORE50 (K50), N3-Reuters-128 (R128), N3-RSS-500 (R500), OKE-15 (O15), and OKE-16 (O16) test sets. Bold indicates the best model. GENRE uses mention dictionaries. The AIT column shows the time in minutes and seconds (m:s) that the systems need to process the whole AIDA test set using an NVIDIA RTX 4090, except for EntQA which does not fit in 24GB of RAM and for which an A100 is used.
To evaluate ReLiK we use the following steps:
-
Download the GERBIL server from here.
-
Start the GERBIL server:
cd gerbil && ./start.sh
- Start the following services:
cd gerbil-SpotWrapNifWS4Test && mvn clean -Dmaven.tomcat.port=1235 tomcat:run
- Start the ReLiK server for GERBIL providing the model name as an argument (e.g.
sapienzanlp/relik-entity-linking-large):
python relik/reader/utils/gerbil.py --relik-model-name sapienzanlp/relik-entity-linking-large
- Open the URL http://localhost:1234/gerbil and:
- Select A2KB as experiment type
- Select "Ma - strong annotation match"
- In the Name field write the name you want to give to the experiment
- In the URI field write: http://localhost:1235/gerbil-spotWrapNifWS4Test/myalgorithm
- Select the datasets (We use AIDA-B, MSNBC, Der, K50, R128, R500, OKE15, OKE16)
- Finally, run experiment
Relation Extraction
The following table shows the results (Micro F1) of ReLiK Large on the NYT dataset:
| Model | NYT | NYT (Pretr) | AIT (m:s) |
|---|---|---|---|
| REBEL | 93.1 | 93.4 | 01:45 |
| UiE | 93.5 | -- | -- |
| USM | 94.0 | 94.1 | -- |
| ReLiKLarge | 95.0 | 94.9 | 00:30 |
To evaluate Relation Extraction we can directly use the reader with the script relik/reader/trainer/predict_re.py, pointing at the file with already retrieved candidates. If you want to use our trained Reader:
python relik/reader/trainer/predict_re.py --model_path sapienzanlp/relik-reader-deberta-v3-large-nyt --data_path /Users/perelluis/Documents/relik/data/debug/test.window.candidates.jsonl --is-eval
Be aware that we compute the threshold for predicting relations based on the development set. To compute it while evaluating you can run the following:
python relik/reader/trainer/predict_re.py --model_path sapienzanlp/relik-reader-deberta-v3-large-nyt --data_path /Users/perelluis/Documents/relik/data/debug/dev.window.candidates.jsonl --is-eval --compute-threshold
๐ฝ Cite this work
If you use any part of this work, please consider citing the paper as follows:
@inproceedings{orlando-etal-2024-relik,
title = "Retrieve, Read and LinK: Fast and Accurate Entity Linking and Relation Extraction on an Academic Budget",
author = "Orlando, Riccardo and Huguet Cabot, Pere-Llu{\'\i}s and Barba, Edoardo and Navigli, Roberto",
booktitle = "Findings of the Association for Computational Linguistics: ACL 2024",
month = aug,
year = "2024",
address = "Bangkok, Thailand",
publisher = "Association for Computational Linguistics",
}
๐ชช License
The data and software are licensed under Creative Commons Attribution-NonCommercial-ShareAlike 4.0.
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
