PYthon Automated Term Extraction

Python implementation of term extraction algorithms such as C-Value, Basic, Combo Basic, Weirdness and Term Extractor using spaCy POS tagging.

If you have a suggestion for another ATE algorithm you would like implemented in this package feel free to file it as an issue with the paper the algorithm is based on.

For ATE packages implemented in Scala and Java, see ATR4S and JATE, respectively.

🎉 Installation

Using pip:

pip install pyate https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-2.2.5/en_core_web_sm-2.2.5.tar.gz

Models

Though this model was originally intended for symbolic AI algorithms (non-machine learning), I realized a spaCy model on term extraction can reach significantly higher performance, and thus decided to include the model here.

For a comparison with the symbolic AI algorithms, see Precision. Note that only the F-Score, accuracy and precision was taken here yet for the model, but for the algorithms the AvP was taken so directly comparing the metrics would not really make sense.

URL	F-Score (%)	Precision (%)	Recall (%)
https://github.com/kevinlu1248/pyate/releases/download/v0.4.2/en_acl_terms_sm-2.0.4.tar.gz	94.71	95.41	94.03

The model was trained and evaluated on the ACL dataset, which is a computer science oriented dataset where the terms are manually picked. This has not yet been tested on other fields yet, however.

This model does not come with PyATE. To install, run

pip install https://github.com/kevinlu1248/pyate/releases/download/v0.4.2/en_acl_terms_sm-2.0.3.tar.gz

To extract terms,

import spacy

nlp = spacy.load("en_acl_terms_sm")
doc = nlp("Hello world, I am a term extraction algorithm.")
print(doc.ents)
"""
(term extraction, algorithm)
"""

🚀 Quickstart

To get started, simply call one of the implemented algorithms. According to Astrakhantsev 2016, combo_basic is the most precise of the five algorithms, though basic and cvalues is not too far behind (see Precision). The same study shows that PU-ATR and KeyConceptRel have higher precision than combo_basic but are not implemented and PU-ATR take significantly more time since it uses machine learning.

from pyate import combo_basic

# source: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1994795/
string = """Central to the development of cancer are genetic changes that endow these “cancer cells” with many of the
hallmarks of cancer, such as self-sufficient growth and resistance to anti-growth and pro-death signals. However, while the
genetic changes that occur within cancer cells themselves, such as activated oncogenes or dysfunctional tumor suppressors,
are responsible for many aspects of cancer development, they are not sufficient. Tumor promotion and progression are
dependent on ancillary processes provided by cells of the tumor environment but that are not necessarily cancerous 
themselves. Inflammation has long been associated with the development of cancer. This review will discuss the reflexive 
relationship between cancer and inflammation with particular focus on how considering the role of inflammation in physiologic 
processes such as the maintenance of tissue homeostasis and repair may provide a logical framework for understanding the U
connection between the inflammatory response and cancer."""

print(combo_basic(string).sort_values(ascending=False))
""" (Output)
dysfunctional tumor                1.443147
tumor suppressors                  1.443147
genetic changes                    1.386294
cancer cells                       1.386294
dysfunctional tumor suppressors    1.298612
logical framework                  0.693147
sufficient growth                  0.693147
death signals                      0.693147
many aspects                       0.693147
inflammatory response              0.693147
tumor promotion                    0.693147
ancillary processes                0.693147
tumor environment                  0.693147
reflexive relationship             0.693147
particular focus                   0.693147
physiologic processes              0.693147
tissue homeostasis                 0.693147
cancer development                 0.693147
dtype: float64
"""

If you would like to add this to a spacy pipeline, simply use add Spacy's add_pipe method.

import spacy
from pyate.term_extraction_pipeline import TermExtractionPipeline

nlp = spacy.load("en_core_web_sm")
nlp.add_pipe(TermExtractionPipeline())
doc = nlp(string)
print(doc._.combo_basic.sort_values(ascending=False).head(5))
""" (Output)
dysfunctional tumor                1.443147
tumor suppressors                  1.443147
genetic changes                    1.386294
cancer cells                       1.386294
dysfunctional tumor suppressors    1.298612
dtype: float64
"""

Also, TermExtractionPipeline.__init__ is defined as follows

__init__(
  self,
  func: Callable[..., pd.Series] = combo_basic,
  *args,
  **kwargs
)

where func is essentially your term extracting algorithm that takes in a corpus (either a string or iterator of strings) and outputs a Pandas Series of term-value pairs of terms and their respective termhoods. func is by default combo_basic. args and kwargs are for you to overide default values for the function, which you can find by running help (might document later on).

Summary of functions

Each of cvalues, basic, combo_basic, weirdness and term_extractor take in a string or an iterator of strings and outputs a Pandas Series of term-value pairs, where higher values indicate higher chance of being a domain specific term. Furthermore, weirdness and term_extractor take a general_corpus key word argument which must be an iterator of strings which defaults to the General Corpus described below.

All functions only take the string of which you would like to extract terms from as the mandatory input (the technical_corpus), as well as other tweakable settings, including general_corpus (contrasting corpus for weirdness and term_extractor), general_corpus_size, verbose (whether to print a progress bar), weights, smoothing, have_single_word (whether to have a single word count as a phrase) and threshold. If you have not read the papers and are unfamiliar with the algorithms, I recommend just using the default settings. Again, use help to find the details regarding each algorithm since they are all different.

General Corpus

Under path/to/site-packages/pyate/default_general_domain.en.csv, there is a general CSV file of a general corpus, specifically, 3000 random sentences from Wikipedia. The source of it can be found at https://www.kaggle.com/mikeortman/wikipedia-sentences. Access it using it using the following after installing pyate.

import pandas as pd
from distutils.sysconfig import get_python_lib  
df = pd.read_csv(get_python_lib() + "/pyate/default_general_domain.en.csv")["SECTION_TEXT"]
print(df.head())
""" (Output)
0    '''Anarchism''' is a political philosophy that...
1    The term ''anarchism'' is a compound word comp...
2    ===Origins===\nWoodcut from a Diggers document...
3    Portrait of philosopher Pierre-Joseph Proudhon...
4    consistent with anarchist values is a controve...
Name: SECTION_TEXT, dtype: object
"""

Other Languages

For switching languages, simply run Term_Extraction.set_language({language}, {model_name}), where model_name defaults to language. For example, Term_Extraction.set_language("it", "it_core_news_sm"}) for Italian. By default, the language is English. So far, only English (en) and Italian (it) are supported.

To add more languages, file an issue with a corpus of at least 3000 paragraphs of a general domain in the desired language (preferably wikipedia) named default_general_domain.{lang}.csv replacing lang with the ISO-639-1 code of the language, or the ISO-639-2 if the language does not have a ISO-639-1 code (can be found at https://www.loc.gov/standards/iso639-2/php/code_list.php). The file format should be of the following form to be parsable by Pandas.

,SECTION_TEXT
0,"{paragraph_0}"
1,"{paragraph_1}"
...

Alternatively, place the file in src/pyate and file a pull request.

🎯 Precision

Here is the average precision of some of the implemented algorithms using the Average Precision (AvP) metric on seven distinct databases, as tested in Astrakhantsev 2016.

🌠 Motivation

This project was planned to be a tool to be connected to a Google Chrome Extension that highlights and defines key terms that the reader probably does not know of. Furthermore, term extraction is an area where there is not a lot of focused research on in comparison to other areas of NLP and especially recently is not viewed to be very practical due to the more general tool of NER tagging. However, modern NER tagging usually incorporates some combination of memorized words and deep learning which are spatially and computationally heavy. Furthermore, to generalize an algorithm to recognize terms to the ever growing areas of medical and AI research, a list of memorized words will not do.

Of the five implemented algorithms, none are expensive, in fact, the bottleneck of the space allocation and computation expense is from the spaCy model and spaCy POS tagging. This is because they mostly rely simply on POS patterns, word frequencies, and the existence of embedded term candidates. For example, the term candidate "breast cancer" implies that "malignant breast cancer" is probably not a term and simply a form of "breast cancer" that is "malignant" (implemented in C-Value).

📌 Todo

Add other languages and data encapsulation for set language
Add automated tests and CI/CD
Add a brief CLI
Make NER version of this using the datasets from the sources
Add PU-ATR algorithm since its precision is a lot higher, though more computationally expensive
Page Rank algorithm
Add sources
Add voting algorithm and capabilities
Optimize perhaps using Cython, however, the bottleneck is POS tagging by Spacy and word counting with Pandas and Numpy, which are already at C-level so this will not help much
Clearer documentation
Allow GPU acceleration with Cupy

📑 Sources

I cannot seem to find the original Basic and Combo Basic papers but I found papers that referenced them. "ATR4S: Toolkit with State-of-the-art Automatic Terms Recognition Methods in Scala" more or less summarizes everything and incorporates several algorithms not in this package.

stelmath/pyate