Skip to content

This repository contains an easy and intuitive approach to few-shot NER using most similar expansion over spaCy embeddings. Now with entity scoring.

License

Notifications You must be signed in to change notification settings

davidberenstein1957/concise-concepts

Repository files navigation

Concise Concepts

When wanting to apply NER to concise concepts, it is really easy to come up with examples, but pretty difficult to train an entire pipeline. Concise Concepts uses few-shot NER based on word embedding similarity to get you going with easy! Now with entity scoring!

Python package Current Release Version pypi Version PyPi downloads Code style: black

Usage

This library defines matching patterns based on the most similar words found in each group, which are used to fill a spaCy EntityRuler. To better understand the rule definition, I recommend playing around with the spaCy Rule-based Matcher Explorer.

Tutorials

The section Matching Pattern Rules expands on the construction, analysis and customization of these matching patterns.

Install

pip install concise-concepts

Quickstart

Take a look at the configuration section for more info.

Spacy Pipeline Component

Note that, custom embedding models are passed via model_path.

import spacy
from spacy import displacy

data = {
    "fruit": ["apple", "pear", "orange"],
    "vegetable": ["broccoli", "spinach", "tomato"],
    "meat": ['beef', 'pork', 'turkey', 'duck']
}

text = """
    Heat the oil in a large pan and add the Onion, celery and carrots.
    Then, cook over a medium–low heat for 10 minutes, or until softened.
    Add the courgette, garlic, red peppers and oregano and cook for 2–3 minutes.
    Later, add some oranges and chickens. """

nlp = spacy.load("en_core_web_md", disable=["ner"])

nlp.add_pipe(
    "concise_concepts",
    config={
        "data": data,
        "ent_score": True,  # Entity Scoring section
        "verbose": True,
        "exclude_pos": ["VERB", "AUX"],
        "exclude_dep": ["DOBJ", "PCOMP"],
        "include_compound_words": False,
        "json_path": "./fruitful_patterns.json",
        "topn": (100,500,300)
    },
)
doc = nlp(text)

options = {
    "colors": {"fruit": "darkorange", "vegetable": "limegreen", "meat": "salmon"},
    "ents": ["fruit", "vegetable", "meat"],
}

ents = doc.ents
for ent in ents:
    new_label = f"{ent.label_} ({ent._.ent_score:.0%})"
    options["colors"][new_label] = options["colors"].get(ent.label_.lower(), None)
    options["ents"].append(new_label)
    ent.label_ = new_label
doc.ents = ents

displacy.render(doc, style="ent", options=options)

Standalone

This might be useful when iterating over few_shot training data when not wanting to reload larger models continuously. Note that, custom embedding models are passed via model.

import gensim
import spacy

from concise_concepts import Conceptualizer

model = gensim.downloader.load("fasttext-wiki-news-subwords-300")
nlp = spacy.load("en_core_web_sm")
data = {
    "disease": ["cancer", "diabetes", "heart disease", "influenza", "pneumonia"],
    "symptom": ["headache", "fever", "cough", "nausea", "vomiting", "diarrhea"],
}
conceptualizer = Conceptualizer(nlp, data, model)
conceptualizer.nlp("I have a headache and a fever.").ents

data = {
    "disease": ["cancer", "diabetes"],
    "symptom": ["headache", "fever"],
}
conceptualizer = Conceptualizer(nlp, data, model)
conceptualizer.nlp("I have a headache and a fever.").ents

Configuration

Matching Pattern Rules

A general introduction about the usage of matching patterns in the usage section.

Customizing Matching Pattern Rules

Even though the baseline parameters provide a decent result, the construction of these matching rules can be customized via the config passed to the spaCy pipeline.

  • exclude_pos: A list of POS tags to be excluded from the rule-based match.
  • exclude_dep: A list of dependencies to be excluded from the rule-based match.
  • include_compound_words: If True, it will include compound words in the entity. For example, if the entity is "New York", it will also include "New York City" as an entity.
  • case_sensitive: Whether to match the case of the words in the text.

Analyze Matching Pattern Rules

To motivate actually looking at the data and support interpretability, the matching patterns that have been generated are stored as ./main_patterns.json. This behavior can be changed by using the json_path variable via the config passed to the spaCy pipeline.

Fuzzy matching using spaczz

  • fuzzy: A boolean value that determines whether to use fuzzy matching
data = {
    "fruit": ["apple", "pear", "orange"],
    "vegetable": ["broccoli", "spinach", "tomato"],
    "meat": ["beef", "pork", "fish", "lamb"]
}

nlp.add_pipe("concise_concepts", config={"data": data, "fuzzy": True})

Most Similar Word Expansion

  • topn: Use a specific number of words to expand over.
data = {
    "fruit": ["apple", "pear", "orange"],
    "vegetable": ["broccoli", "spinach", "tomato"],
    "meat": ["beef", "pork", "fish", "lamb"]
}

topn = [50, 50, 150]

assert len(topn) == len

nlp.add_pipe("concise_concepts", config={"data": data, "topn": topn})

Entity Scoring

  • ent_score: Use embedding based word similarity to score entities against their groups
import spacy

data = {
    "ORG": ["Google", "Apple", "Amazon"],
    "GPE": ["Netherlands", "France", "China"],
}

text = """Sony was founded in Japan."""

nlp = spacy.load("en_core_web_lg")
nlp.add_pipe("concise_concepts", config={"data": data, "ent_score": True, "case_sensitive": True})
doc = nlp(text)

print([(ent.text, ent.label_, ent._.ent_score) for ent in doc.ents])
# output
#
# [('Sony', 'ORG', 0.5207586), ('Japan', 'GPE', 0.7371268)]

Custom Embedding Models

  • model_path: Use custom sense2vec.Sense2Vec, gensim.Word2vec gensim.FastText, or gensim.KeyedVectors, or a pretrained model from gensim library or a custom model path. For using a sense2vec.Sense2Vec take a look here.
  • model: within standalone usage, it is possible to pass these models directly.
data = {
    "fruit": ["apple", "pear", "orange"],
    "vegetable": ["broccoli", "spinach", "tomato"],
    "meat": ["beef", "pork", "fish", "lamb"]
}

# model from https://radimrehurek.com/gensim/downloader.html or path to local file
model_path = "glove-wiki-gigaword-300"

nlp.add_pipe("concise_concepts", config={"data": data, "model_path": model_path})