/zshot

Zero and Few shot named entity & relationships recognition

Primary LanguagePythonMIT LicenseMIT

Zshot

Zero and Few shot named entity & relationships recognition

Tutorials Build Build

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

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

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.
  • blink - Required if you want to use Blink for linking to Wikipedia pages.

Installation

$ pip install zshot

---> 100%

Examples

Example Notebook
Installation and Visualization Open In Colab

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 4 different mentions extractors supported, 2 of them are based on SpaCy, and 2 of them are based on Flair. The two different versions for each library 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 4 linkers available currently, 2 of them are end-to-end and 2 are not. Let's start with those thar are not end-to-end:

Linker Name end-to-end Source Code Paper
Blink X Source Code Paper
GENRE X Source Code Paper
SMXM Source Code Paper
TARS Source Code Paper

Example: Zero-Shot Entity Recognition

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(s) and split(s) to evaluate. It will return a str containing a table with the results of the evaluation. For instance the evaluation of the ZShot custom component implemented above would be:

from zshot.evaluation.zshot_evaluate import evaluate
from datasets import Split

evaluation = evaluate(new_nlp, datasets="ontonotes", 
                      splits=[Split.VALIDATION])
print(evaluation)