knowledgator/gliner-multitask-v1.0

🚀 Meet the first multi-task prompt-tunable GLiNER model 🚀

GLiNER-Multitask is a model designed to extract various pieces of information from plain text based on a user-provided custom prompt. This versatile model leverages a bidirectional transformer encoder, similar to BERT, which ensures both high generalization and compute efficiency despite its compact size.

The gliner-multitask-v1.0 variant achieves state-of-the-art performance on NER zero-shot benchmarks, demonstrating its robustness and flexibility. It excels not only in named entity recognition but also in handling various other information extraction tasks, making it a powerful tool for diverse natural language processing applications.

Supported tasks:

• Named Entity Recognition (NER): Identifies and categorizes entities such as names, organizations, dates, and other specific items in the text.
• Relation Extraction: Detects and classifies relationships between entities within the text.
• Summarization: Extract the most important sentences that summarize the input text, capturing the essential information.
• Sentiment Extraction: Identify parts of the text that signalize a positive, negative, or neutral sentiment;
• Key-Phrase Extraction: Identifies and extracts important phrases and keywords from the text.
• Question-answering: Finding an answer in the text given a question;
• Open Information Extraction: Extracts pieces of text given an open prompt from a user, for example, product description extraction;
• Text classification: Classifying text by matching labels specified in the prompt;

Installation

To use this model, you must install the GLiNER Python library:

pip install gliner

Once you've downloaded the GLiNER library, you can import the GLiNER class. You can then load this model using GLiNER.from_pretrained.

How to use for NER:

from gliner import GLiNER

model = GLiNER.from_pretrained("knowledgator/gliner-multitask-v1.0")

text = """
Microsoft was founded by Bill Gates and Paul Allen on April 4, 1975 to develop and sell BASIC interpreters for the Altair 8800. During his career at Microsoft, Gates held the positions of chairman, chief executive officer, president and chief software architect, while also being the largest individual shareholder until May 2014.
"""

labels = ["founder", "computer", "software", "position", "date"]

entities = model.predict_entities(text, labels)

for entity in entities:
    print(entity["text"], "=>", entity["label"])

Performance:

Model	Dataset	Precision	Recall	F1 Score	F1 Score (Decimal)
knowledgator/gliner-multitask-v0.5	CrossNER_AI	51.00%	51.11%	51.05%	0.5105
	CrossNER_literature	72.65%	65.62%	68.96%	0.6896
	CrossNER_music	74.91%	73.70%	74.30%	0.7430
	CrossNER_politics	78.84%	77.71%	78.27%	0.7827
	CrossNER_science	69.20%	65.48%	67.29%	0.6729
	mit-movie	61.29%	52.59%	56.60%	0.5660
	mit-restaurant	50.65%	38.13%	43.51%	0.4351
	Average				0.6276
knowledgator/gliner-multitask-v1.0	CrossNER_AI	67.15%	56.10%	61.13%	0.6113
	CrossNER_literature	71.60%	64.74%	68.00%	0.6800
	CrossNER_music	73.57%	69.29%	71.36%	0.7136
	CrossNER_politics	77.54%	76.52%	77.03%	0.7703
	CrossNER_science	74.54%	66.00%	70.01%	0.7001
	mit-movie	61.86%	42.02%	50.04%	0.5004
	mit-restaurant	58.87%	36.67%	45.19%	0.4519
	Average				0.6325
knowledgator/gliner-llama-multitask-1B-v1.0	CrossNER_AI	63.24%	55.60%	59.17%	0.5917
	CrossNER_literature	69.74%	60.10%	64.56%	0.6456
	CrossNER_music	74.03%	67.22%	70.46%	0.7046
	CrossNER_politics	76.96%	71.64%	74.20%	0.7420
	CrossNER_science	73.79%	63.73%	68.39%	0.6839
	mit-movie	56.89%	46.70%	51.30%	0.5130
	mit-restaurant	48.45%	38.13%	42.67%	0.4267
	Average				0.6153

---

How to use for relation extraction:

text = """
Microsoft was founded by Bill Gates and Paul Allen on April 4, 1975 to develop and sell BASIC interpreters for the Altair 8800. During his career at Microsoft, Gates held the positions of chairman, chief executive officer, president and chief software architect, while also being the largest individual shareholder until May 2014.
"""

labels = ["Microsoft <> founder", "Microsoft <> inception date", "Bill Gates <> held position"]

entities = model.predict_entities(text, labels)

for entity in entities:
    print(entity["label"], "=>", entity["text"])

Construct relations extraction pipeline with utca

First of all, we need import neccessary components of the library and initalize predictor - GLiNER model and construct pipeline that combines NER and realtions extraction:

from utca.core import RenameAttribute
from utca.implementation.predictors import (
    GLiNERPredictor,
    GLiNERPredictorConfig
)
from utca.implementation.tasks import (
    GLiNER,
    GLiNERPreprocessor,
    GLiNERRelationExtraction,
    GLiNERRelationExtractionPreprocessor,
)

predictor = GLiNERPredictor( # Predictor manages the model that will be used by tasks
    GLiNERPredictorConfig(
        model_name = "knowledgator/gliner-multitask-v1.0", # Model to use
        device = "cuda:0", # Device to use
    )
)

pipe = (
    GLiNER( # GLiNER task produces classified entities that will be at the "output" key.
        predictor=predictor,
        preprocess=GLiNERPreprocessor(threshold=0.7) # Entities threshold
    ) 
    | RenameAttribute("output", "entities") # Rename output entities from GLiNER task to use them as inputs in GLiNERRelationExtraction
    | GLiNERRelationExtraction( # GLiNERRelationExtraction is used for relation extraction.
        predictor=predictor,
        preprocess=(
            GLiNERPreprocessor(threshold=0.5) # Relations threshold
            | GLiNERRelationExtractionPreprocessor()
        )
    )
)

To run pipeline we need to specify entity types and relations with their parameters:

r = pipe.run({
    "text": text, # Text to process
    "labels": ["organisation", "founder", "position", "date"],
    "relations": [{ # Relation parameters
        "relation": "founder", # Relation label. Required parameter.
        "pairs_filter": [("organisation", "founder")], # Optional parameter. It specifies possible members of relations by their entity labels.
        "distance_threshold": 100, # Optional parameter. It specifies the max distance between spans in the text (i.e., the end of the span that is closer to the start of the text and the start of the next one).
    }, {
        "relation": "inception date",
        "pairs_filter": [("organisation", "date")],
    }, {
        "relation": "held position",
        "pairs_filter": [("founder", "position")],
    }]
})

print(r["output"])

Performance:

Model	Dataset	Precision	Recall	F1 Score
knowledgator/gliner-llama-multitask-1B-v1.0	CrossRe	0.606472	0.511444	0.554919
	DocRed	0.707483	0.589355	0.643039
knowledgator/gliner-multitask-v0.5	CrossRe	0.585319	0.800176	0.676088
	DocRed	0.713392	0.772826	0.74192
knowledgator/gliner-multitask-v1.0	CrossRe	0.760653	0.738556	0.749442
	DocRed	0.770644	0.761373	0.76598

---

How to use for open information extraction:

prompt = """Find all positive aspects about the product:\n"""
text = """
I recently purchased the Sony WH-1000XM4 Wireless Noise-Canceling Headphones from Amazon and I must say, I'm thoroughly impressed. The package arrived in New York within 2 days, thanks to Amazon Prime's expedited shipping.

The headphones themselves are remarkable. The noise-canceling feature works like a charm in the bustling city environment, and the 30-hour battery life means I don't have to charge them every day. Connecting them to my Samsung Galaxy S21 was a breeze, and the sound quality is second to none.

I also appreciated the customer service from Amazon when I had a question about the warranty. They responded within an hour and provided all the information I needed.

However, the headphones did not come with a hard case, which was listed in the product description. I contacted Amazon, and they offered a 10% discount on my next purchase as an apology.

Overall, I'd give these headphones a 4.5/5 rating and highly recommend them to anyone looking for top-notch quality in both product and service.
"""

input_ = prompt+text

labels = ["match"]

matches = model.predict_entities(input_, labels)

for match in matches:
    print(match["text"], "=>", match["score"])

Performance:

Dataset: WiRe57_343-manual-oie

Model	Precision	Recall	F1 Score
knowledgator/gliner-llama-multitask-1B-v1.0	0.9047	0.2794	0.4269
knowledgator/gliner-multitask-v0.5	0.9278	0.2779	0.4287
knowledgator/gliner-multitask-v1.0	0.8775	0.2733	0.4168

---

How to use for question-answering:

question = "Who was the CEO of Microsoft?"
text = """
Microsoft was founded by Bill Gates and Paul Allen on April 4, 1975, to develop and sell BASIC interpreters for the Altair 8800. During his career at Microsoft, Gates held the positions of chairman, chief executive officer, president and chief software architect, while also being the largest individual shareholder until May 2014.
"""

labels = ["answer"]

input_ = question+text
answers = model.predict_entities(input_, labels)

for answer in answers:
    print(answer["text"], "=>", answer["score"])

Performance:

Dataset: SQuAD 2.0

Model	Precision	Recall	F1 Score
knowledgator/gliner-llama-multitask-1B-v1.0	0.578296	0.795821	0.669841
knowledgator/gliner-multitask-v0.5	0.429213	0.94378	0.590072
knowledgator/gliner-multitask-v1.0	0.601354	0.874784	0.712745

---

How to use for summarization:

With threshold parameters, you can control how much information you want to extract.

prompt = "Summarize the given text, highlighting the most important information:\n"

text = """
Several studies have reported its pharmacological activities, including anti-inflammatory, antimicrobial, and antitumoral effects.
The effect of E-anethole was studied in the osteosarcoma MG-63 cell line, and the antiproliferative activity was evaluated by an MTT assay.
It showed a GI50 value of 60.25 μM with apoptosis induction through the mitochondrial-mediated pathway. Additionally, it induced cell cycle arrest at the G0/G1 phase, up-regulated the expression of p53, caspase-3, and caspase-9, and down-regulated Bcl-xL expression.
Moreover, the antitumoral activity of anethole was assessed against oral tumor Ca9-22 cells, and the cytotoxic effects were evaluated by MTT and LDH assays.
It demonstrated a LD50 value of 8 μM, and cellular proliferation was 42.7% and 5.2% at anethole concentrations of 3 μM and 30 μM, respectively.
It was reported that it could selectively and in a dose-dependent manner decrease cell proliferation and induce apoptosis, as well as induce autophagy, decrease ROS production, and increase glutathione activity. The cytotoxic effect was mediated through NF-kB, MAP kinases, Wnt, caspase-3 and -9, and PARP1 pathways. Additionally, treatment with anethole inhibited cyclin D1 oncogene expression, increased cyclin-dependent kinase inhibitor p21WAF1, up-regulated p53 expression, and inhibited the EMT markers.
"""

labels = ["summary"]

input_ = prompt+text

threshold = 0.1
summaries = model.predict_entities(input_, labels, threshold=threshold)

for summary in summaries:
    print(summary["text"], "=>", summary["score"])

---

How to use for text classification:

With threshold parameters, you can control recall and precision of text classification.

prompt = "Classify text into the following classes: positive review, negative review"

text = """
"I recently purchased the Sony WH-1000XM4 Wireless Noise-Canceling Headphones from Amazon and I must say, I'm thoroughly impressed. The package arrived in New York within 2 days, thanks to Amazon Prime's expedited shipping.
"""

labels = ["match"]

input_ = prompt+text

threshold = 0.5
classes = model.predict_entities(input_, labels, threshold=threshold)

for label in classes:
    print(label["text"], "=>", label["score"])

Performance:

Model Name	Dataset	Micro F1 Score
knowledgator/gliner-multitask-v1.0	Emotion	0.322
	AG News	0.7436
	IMDb	0.7907
knowledgator/gliner-llama-multitask-1B-v1.0	Emotion	0.3475
	AG News	0.7436
	IMDb	0.7907

---

Extensive NER Benchmarks:

Our multitask model demonstrates comparable performance on different zero-shot benchmarks to dedicated models to NER task (all labels were lowecased in this testing):

Dataset	Precision	Recall	F1 Score	F1 Score (Decimal)
ACE 2004	53.25%	23.20%	32.32%	0.3232
ACE 2005	43.25%	18.00%	25.42%	0.2542
AnatEM	51.75%	25.98%	34.59%	0.3459
Broad Tweet Corpus	69.54%	72.50%	70.99%	0.7099
CoNLL 2003	68.33%	68.43%	68.38%	0.6838
CrossNER_AI	67.15%	56.10%	61.13%	0.6113
CrossNER_literature	71.60%	64.74%	68.00%	0.6800
CrossNER_music	73.57%	69.29%	71.36%	0.7136
CrossNER_politics	77.54%	76.52%	77.03%	0.7703
CrossNER_science	74.54%	66.00%	70.01%	0.7001
FabNER	69.28%	62.62%	65.78%	0.6578
FindVehicle	49.75%	51.25%	50.49%	0.5049
GENIA_NER	60.98%	46.91%	53.03%	0.5303
HarveyNER	24.27%	35.66%	28.88%	0.2888
MultiNERD	54.33%	89.34%	67.57%	0.6757
Ontonotes	27.26%	36.64%	31.26%	0.3126
PolyglotNER	33.54%	64.29%	44.08%	0.4408
TweetNER7	44.77%	38.67%	41.50%	0.4150
WikiANN en	56.33%	57.09%	56.71%	0.5671
WikiNeural	71.70%	86.60%	78.45%	0.7845
bc2gm	64.71%	51.68%	57.47%	0.5747
bc4chemd	69.24%	50.08%	58.12%	0.5812
bc5cdr	79.22%	69.19%	73.87%	0.7387
mit-movie	61.86%	42.02%	50.04%	0.5004
mit-restaurant	58.87%	36.67%	45.19%	0.4519
ncbi	68.72%	54.86%	61.01%	0.6101

Join Our Discord

Connect with our community on Discord for news, support, and discussion about our models. Join Discord.

Citation:

@misc{stepanov2024gliner,
      title={GLiNER multi-task: Generalist Lightweight Model for Various Information Extraction Tasks}, 
      author={Ihor Stepanov and Mykhailo Shtopko},
      year={2024},
      eprint={2406.12925},
      archivePrefix={arXiv},
      primaryClass={id='cs.LG' full_name='Machine Learning' is_active=True alt_name=None in_archive='cs' is_general=False description='Papers on all aspects of machine learning research (supervised, unsupervised, reinforcement learning, bandit problems, and so on) including also robustness, explanation, fairness, and methodology. cs.LG is also an appropriate primary category for applications of machine learning methods.'}
}