Zshot

Zero and Few shot named entity & relationships recognition

Documentation: https://ibm.github.io/zshot

Source Code: https://github.com/IBM/zshot

Paper: https://aclanthology.org/2023.acl-demo.34/

Zshot is a highly customisable framework for performing Zero and Few shot named entity recognition.

Can be used to perform:

• Mentions extraction: Identify globally relevant mentions or mentions relevant for a given domain
• Wikification: The task of linking textual mentions to entities in Wikipedia
• Zero and Few Shot named entity recognition: using language description perform NER to generalize to unseen domains
• Zero and Few Shot named relationship recognition
• Visualization: Zero-shot NER and RE extraction

Requirements

• Python 3.6+

• spacy - Zshot rely on Spacy for pipelining and visualization

• torch - PyTorch is required to run pytorch models.

• transformers - Required for pre-trained language models.

• evaluate - Required for evaluation.

• datasets - Required to evaluate over datasets (e.g.: OntoNotes).

Optional Dependencies

• flair - Required if you want to use Flair mentions extractor and for TARS linker and TARS Mentions Extractor.
• blink - Required if you want to use Blink for linking to Wikipedia pages.
• gliner - Required if you want to use GLiNER Linker or GLiNER Mentions Extractor.

Installation

Examples

Example	Notebook
Installation and Visualization
Knowledge Extractor
Wikification
Custom Components
Evaluation

Zshot Approach

ZShot contains two different components, the mentions extractor and the linker.

Mentions Extractor

The mentions extractor will detect the possible entities (a.k.a. mentions), that will be then linked to a data source (e.g.: Wikidata) by the linker.

Currently, there are 7 different mentions extractors supported, SMXM, TARS, GLiNER, 2 based on SpaCy, and 2 that are based on Flair. The two different versions for SpaCy and Flair are similar, one is based on Named Entity Recognition and Classification (NERC) and the other one is based on the linguistics (i.e.: using Part Of the Speech tagging (PoS) and Dependency Parsing(DP)).

The NERC approach will use NERC models to detect all the entities that have to be linked. This approach depends on the model that is being used, and the entities the model has been trained on, so depending on the use case and the target entities it may be not the best approach, as the entities may be not recognized by the NERC model and thus won't be linked.

The linguistic approach relies on the idea that mentions will usually be a syntagma or a noun. Therefore, this approach detects nouns that are included in a syntagma and that act like objects, subjects, etc. This approach do not depend on the model (although the performance does), but a noun in a text should be always a noun, it doesn't depend on the dataset the model has been trained on.

Linker

The linker will link the detected entities to a existing set of labels. Some of the linkers, however, are end-to-end, i.e. they don't need the mentions extractor, as they detect and link the entities at the same time.

Again, there are 5 linkers available currently, 3 of them are end-to-end and 2 are not.

Relations Extractor

The relations extractor will extract relations among different entities previously extracted by a linker..

Currently, the is only one Relation Extractor available:

• ZS-Bert
  • Paper
  • Source Code

Knowledge Extractor

The knowledge extractor will perform at the same time the extraction and classification of named entities and the extraction of relations among them. The pipeline with this component doesn't need any mentions extractor, linker or relation extractor to work.

Currently, the is only one Knowledge Extractor available:

• KnowGL

  • Rossiello et al. (AAAI 2023)
  • Mihindukulasooriya et al. (ISWC 2022)

How to use it

• Install requirements: pip install -r requirements.txt
• Install a spacy pipeline to use it for mentions extraction: python -m spacy download en_core_web_sm
• Create a file main.py with the pipeline configuration and entities definition (Wikipedia abstract are usually a good starting point for descriptions):

import spacy

from zshot import PipelineConfig, displacy
from zshot.linker import LinkerRegen
from zshot.mentions_extractor import MentionsExtractorSpacy
from zshot.utils.data_models import Entity

nlp = spacy.load("en_core_web_sm")
nlp_config = PipelineConfig(
    mentions_extractor=MentionsExtractorSpacy(),
    linker=LinkerRegen(),
    entities=[
        Entity(name="Paris",
               description="Paris is located in northern central France, in a north-bending arc of the river Seine"),
        Entity(name="IBM",
               description="International Business Machines Corporation (IBM) is an American multinational technology corporation headquartered in Armonk, New York"),
        Entity(name="New York", description="New York is a city in U.S. state"),
        Entity(name="Florida", description="southeasternmost U.S. state"),
        Entity(name="American",
               description="American, something of, from, or related to the United States of America, commonly known as the United States or America"),
        Entity(name="Chemical formula",
               description="In chemistry, a chemical formula is a way of presenting information about the chemical proportions of atoms that constitute a particular chemical compound or molecule"),
        Entity(name="Acetamide",
               description="Acetamide (systematic name: ethanamide) is an organic compound with the formula CH3CONH2. It is the simplest amide derived from acetic acid. It finds some use as a plasticizer and as an industrial solvent."),
        Entity(name="Armonk",
               description="Armonk is a hamlet and census-designated place (CDP) in the town of North Castle, located in Westchester County, New York, United States."),
        Entity(name="Acetic Acid",
               description="Acetic acid, systematically named ethanoic acid, is an acidic, colourless liquid and organic compound with the chemical formula CH3COOH"),
        Entity(name="Industrial solvent",
               description="Acetamide (systematic name: ethanamide) is an organic compound with the formula CH3CONH2. It is the simplest amide derived from acetic acid. It finds some use as a plasticizer and as an industrial solvent."),
    ]
)
nlp.add_pipe("zshot", config=nlp_config, last=True)

text = "International Business Machines Corporation (IBM) is an American multinational technology corporation" \
       " headquartered in Armonk, New York, with operations in over 171 countries."

doc = nlp(text)
displacy.serve(doc, style="ent")

Run it

Run with

$ python main.py

Using the 'ent' visualizer
Serving on http://0.0.0.0:5000 ...

The script will annotate the text using Zshot and use Displacy for visualising the annotations

Check it

Open your browser at http://127.0.0.1:5000 .

You will see the annotated sentence:

How to create a custom component

If you want to implement your own mentions_extractor or linker and use it with ZShot you can do it. To make it easier for the user to implement a new component, some base classes are provided that you have to extend with your code.

It is as simple as create a new class extending the base class (MentionsExtractor or Linker). You will have to implement the predict method, which will receive the SpaCy Documents and will return a list of zshot.utils.data_models.Span for each document.

This is a simple mentions_extractor that will extract as mentions all words that contain the letter s:

from typing import Iterable
import spacy
from spacy.tokens import Doc
from zshot import PipelineConfig
from zshot.utils.data_models import Span
from zshot.mentions_extractor import MentionsExtractor

class SimpleMentionExtractor(MentionsExtractor):
    def predict(self, docs: Iterable[Doc], batch_size=None):
        spans = [[Span(tok.idx, tok.idx + len(tok)) for tok in doc if "s" in tok.text] for doc in docs]
        return spans

new_nlp = spacy.load("en_core_web_sm")

config = PipelineConfig(
    mentions_extractor=SimpleMentionExtractor()
)
new_nlp.add_pipe("zshot", config=config, last=True)
text_acetamide = "CH2O2 is a chemical compound similar to Acetamide used in International Business " \
        "Machines Corporation (IBM)."

doc = new_nlp(text_acetamide)
print(doc._.mentions)

>>> [is, similar, used, Business, Machines, materials]

How to evaluate ZShot

Evaluation is an important process to keep improving the performance of the models, that's why ZShot allows to evaluate the component with two predefined datasets: OntoNotes and MedMentions, in a Zero-Shot version in which the entities of the test and validation splits don't appear in the train set.

The package evaluation contains all the functionalities to evaluate the ZShot components. The main function is zshot.evaluation.zshot_evaluate.evaluate, that will take as input the SpaCy nlp model and the dataset to evaluate. It will return a str containing a table with the results of the evaluation. For instance the evaluation of the TARS linker in ZShot for the Ontonotes validation set would be:

import spacy

from zshot import PipelineConfig
from zshot.linker import LinkerTARS
from zshot.evaluation.dataset import load_ontonotes_zs
from zshot.evaluation.zshot_evaluate import evaluate, prettify_evaluate_report
from zshot.evaluation.metrics.seqeval.seqeval import Seqeval

ontonotes_zs = load_ontonotes_zs('validation')

nlp = spacy.blank("en")
nlp_config = PipelineConfig(
    linker=LinkerTARS(),
    entities=ontonotes_zs.entities
)

nlp.add_pipe("zshot", config=nlp_config, last=True)

evaluation = evaluate(nlp, ontonotes_zs, metric=Seqeval())
prettify_evaluate_report(evaluation)

Citation

@inproceedings{picco-etal-2023-zshot,
    title = "Zshot: An Open-source Framework for Zero-Shot Named Entity Recognition and Relation Extraction",
    author = "Picco, Gabriele  and
      Martinez Galindo, Marcos  and
      Purpura, Alberto  and
      Fuchs, Leopold  and
      Lopez, Vanessa  and
      Hoang, Thanh Lam",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-demo.34",
    doi = "10.18653/v1/2023.acl-demo.34",
    pages = "357--368",
    abstract = "The Zero-Shot Learning (ZSL) task pertains to the identification of entities or relations in texts that were not seen during training. ZSL has emerged as a critical research area due to the scarcity of labeled data in specific domains, and its applications have grown significantly in recent years. With the advent of large pretrained language models, several novel methods have been proposed, resulting in substantial improvements in ZSL performance. There is a growing demand, both in the research community and industry, for a comprehensive ZSL framework that facilitates the development and accessibility of the latest methods and pretrained models.In this study, we propose a novel ZSL framework called Zshot that aims to address the aforementioned challenges. Our primary objective is to provide a platform that allows researchers to compare different state-of-the-art ZSL methods with standard benchmark datasets. Additionally, we have designed our framework to support the industry with readily available APIs for production under the standard SpaCy NLP pipeline. Our API is extendible and evaluable, moreover, we include numerous enhancements such as boosting the accuracy with pipeline ensembling and visualization utilities available as a SpaCy extension.",
}