Tokenizing CORD-19 with NLTK

Connor Shorten
5 min readNov 27, 2020

Scientific overload is one of the toughest challenges facing scientists today. As Machine Learning researchers, we constantly complain about the fast pace of arxiv uploads and praise organization tools like Arxiv Sanity Preserver. The scientific response to COVID-19 is another example of information overload. The CORD-19 dataset documents over 100K papers containing relevant information. No single or group of human beings could be expected to interpret this amount of information.

We need better search engines for Scientific Papers. Deep Learning powered search engines, question answering systems, or even chatbots and summarizers seem possible. We have data, and we have a lot of great software to work on this. I have been really interested in this application area and hope this collection of notebooks will inspire others to explore the CORD-19 dataset and Scientific NLP.

This tutorial is a quick overview on the basics of tokenization with NLTK:

Example of tokens mapped to their index positions. Clearly we have problems from the NLTK tokenization such as the “data;” token. I will build on this notebook with more data cleaning in the future to avoid these problems. For now, I think this is a solid tutorial on the basics of building token to index lists.

Colab Notebook:

NLTK Tokenizer

The NLTK tokenizer is pretty naive compared to things like Byte-Pair encoding or Subword tokenization. However, it gets the job done and is a great baseline for getting started.

import nltk'punkt');
from nltk.tokenize import word_tokenize
text = "Exploring the CORD19 Dataset"
# ['Exploring', 'the', 'CORD19', 'Dataset']

Single sentence speed test:

text = "Exploring the CORD19 Dataset"
# CPU times: user 279 µs, sys: 0ns, total: 279 µs
# Wall time: 293 µs

Longer test — 100 sentences with 10 words

# CPU times: user 11.8 ms, sys: 0ns, total: 11.8ms
# Wall time: 11.7 ms

Top-K Dictionary

  • The first step in Tokenization is to build a dictionary that maps from a token to its index in the dictionary, e.g. “hello” → 815

Naive strategy — Top-K Frequency Dictionary

  • Loop through the text to count the 30,000 most frequently occurring words
  • Assign each of these top 29,999 words an index mapping their top-K rank. If “the” is the 15th most frequently occurring token… it is mapped to index 15.
  • All words outside of the top-K will be mapped to 30_000
freq_counter = {} # freq short for frequency...
for i in df.Sequence: # Sequence is a DataFrame column containing paragraphs
tok_seq = word_tokenize(i)
for tok in tok_seq:
if tok in freq_counter.keys():
freq_counter[tok] += 1
freq_counter[tok] = 1
# CPU times: user 5min 5s, sys: 770 ms, total: 5min 6s
# Wall time: 5min 7s

Sort the Dictionary according to the top-K counts

import operator
sorted_dict = sorted(freq_counter.items(), key=operator.itemgetter(1))
sorted_dict[-2:]# [('the', 1567251), (',', 1867586)]

Major problem with this tokenization strategy: There are 656,159 unique tokens but we are going to cut our vocabulary to the top 30,000 most frequently occurring tokens. The second line of code is visualizing the first 5 tokens that miss the cut. We see that we miss out on tokens like “Scintillation” or “Proteintech”, we are losing a good deal of information with this truncation, but the bigger the vocabulary, the more computationally expensive our embedding table becomes.

print(len(sorted_dict))print(sorted_dict[-30_000:][:5]# 656159# [('scintillation', 41),
('inositol', 41),
('Proteintech', 41),
('stillbirth', 41),
('colder', 41)]

Continuing despite this problem…

Only select the top 29,999 most frequently occurring tokens

top_K_list = sorted_dict[-29_999:] # 29999 -> 29_999... much more readable

Construct the Token -> Index; Index -> Token dictionaries

top_K_token_to_index_dict = {}
top_K_index_to_token_dict = {}
# ^ verbose naming, but hopefully more clear for sake of tutorialfor i in range(len(top_K_list)):
top_K_token_index_dict[top_K_list[i][0]] = counter
top_K_index_token_dict[counter] = top_K_list[i][0]
counter += 1
top_K_index_token_dict[30_000] = "Unknown"

Save the Dictionaries (Token -> Index; Index -> Token)

import json# Token -> Index
token_index_dict_write = json.dumps(top_K_token_index_dict)
f = open("token_index_dict.json", "w")
# Index -> Token
index_token_dict_write = json.dumps(top_K_index_token_dict)
f = open("index_token_dict.json", "w")

How to load this dictionary

f = open("token_index_dict.json", "r")
dict_text = f.readlines()[0]
token_index_dict = json.loads(dict_text)

Text -> Index Mapping

def text_to_index(seq, token_index_dict):
idx_lst = []
tok_lst = word_tokenize(seq)
for tok in tok_lst:
if tok not in token_index_dict.keys():
return idx_lst


sentence = "hello how are you doing"
text_to_index(sentence, top_K_token_index_dict)
# [30_000, 29675, 29978, 29143, 25834]

Build Index Lists

These index lists are the input to our Neural Network

def build_index_lists(df, text_col_name, text_index_dict):
index_lists = []
for seq in df[text_col_name]:
seq = seq.split(' ')
new_index_list = []
for tok in seq:
if tok in text_index_dict.keys():
return index_lists

Sanity Check of our index lists

Padding or Truncating Sequences to Length k (k=128 in this case)

def pad_to_length_k(org_index_lists, k):
index_lists = org_index_lists
for seq_list = index_lists:
while (len(seq_list) > k):
while (len(seq_list) < k):
return index_lists

Add Index Lists to DataFrame

index_lists = pad_to_length_k(index_lists, 128)
df["Index_Lists"] = index_lists

Save for use later on

df.to_csv('IdxLists_Pdf_Json_1.csv', index=False)

Tokenization Complete: What we need for Downstream Applications

Load the DataFrame with the Index Lists

from google.colab import files

Load Dataframe

df = pd.read_csv('IdxLists_Pdf_Json_1.csv')

Load Text -> Index Dict

f = open("token_index_dict.json", "r")
dict_text = f.readlines()[0]
token_index_dict = json.loads(dict_text)

Load Index -> Text Dict

f = open("index_token_dict.json", "r")
dict_text = f.readlines()[0]
index_token_dict = json.loads(dict_text)

Load Text_to_Index and Index_to_Text functions


git clone

Load and test functions

import sys, os
sys.path.append(os.getcwd() + '/CORD-19-Mining/')
# note, I'll probably group this in a /utils folder soon
# so also try " + '/CORD-19-Mining/utils/') " if it doesn't work
from tokenization import text_to_index, index_to_text


text_to_index("hello how are you", token_index_dict)


This notebook has shown you how to tokenize text stored as rows in DataFrame column with the NLTK tokenizer. This was done by assigning indexes based on the top 30K most frequently occurring words. I expect this tokenization to be too naive for applications like information retrieval, question answering, or summarization. However, I think this is a solid place to get started and at least we have the index lists to build the models and get that half of the pipeline working properly.

Thanks for reading! If you are interested in Deep Learning and AI please subscribe to my YouTube channel — Henry AI Labs. I will be uploading a coding series on mining the CORD-19 dataset soon (probably between March-April 2021).