The corpushash library enables the performance of common NLP tasks on sensitive documents without disclosing their contents. This is done by hashing every token in the corpus along with a salt.

1. the client --- in her own secure environment --- takes her classified documents and does the linguistic pre-processing (removal of stopwords, tokenization, etc.)

2. the client then hashes the tokens: she creates random salts for every different token and hashes the concatenation of token salt. the client keeps a dictionary of the salts used for every token.

3. the client sends the hashed tokens to the analyst for linguistic processing. as there is a biunivocal correspondence between every hashed token and plaintext token, NLP can occur in the same way as with any plaintext corpus.

4. once the NLP is over, the analyst sends the results to the client, who then uses the dictionary to recover the plaintext tokens and thus the results' meaning.

The library requires as input:

• a tokenized corpus as a nested list, whose elements are themselves nested lists of the tokens of each document in the corpus.

  each list corresponds to a document structure: its chapters, paragraphs, sentences. you decide how the nested list is to be created or structured, as long as the input is a nested list with strings as their bottom-most elements..

• corpus_path, a path to a directory where the output files are to be stored.

The output includes:

• a .json file for every document in the corpus, named sequentially as positive integers, e.g., the first document being 0.json, stored in corpus_path/public/$(timestamp-of-hash)/.
• two .json dictionaries stored in corpus_path/private. they are used to decode the .json files or the NLP results.

the package demands python ⩾3, but has no external dependencies.

using pip:

pip3 install corpushash

or from repository (most up-to-date):

pip3 install git https://github.com/NAMD/corpushash.git

or manually:

git clone https://github.com/NAMD/corpushash.git
cd corpushash
python3 setup.py install

this will hash each word in the first four verses of the zen of python to the same .json document:

import corpushash as ch
example_corpus = [[
                 ['Beautiful', 'is', 'better', 'than', 'ugly'],
                 ['Explicit', 'is', 'better', 'than', 'implicit'],
                 ['Simple', 'is', 'better', 'than', 'complex'],
                 ['Complex', 'is', 'better', 'than', 'complicated'],
                 ['Flat', 'is', 'better', 'than', 'nested']
                 ]]
hashed_corpus = ch.CorpusHash(example_corpus, 'output_directory')

this will hash each word in the first four verses of the zen of python to four different .json documents, as if they were different documents:

import corpushash as ch
example_corpus = [
                 ['Beautiful', 'is', 'better', 'than', 'ugly'],
                 ['Explicit', 'is', 'better', 'than', 'implicit'],
                 ['Simple', 'is', 'better', 'than', 'complex'],
                 ['Complex', 'is', 'better', 'than', 'complicated'],
                 ['Flat', 'is', 'better', 'than', 'nested']
                 ]
hashed_corpus = ch.CorpusHash(example_corpus, 'output_directory')

so be careful when constructing your nested lists! check the tutorial at notebooks/tutorial.ipynb.

• probability of collision is extremely low (check the preprint), but still we check for them, so they are not an issue.
• hashing the tokens is not the same as encrypting them. as the same token always maps to the same hash, the resulting hashed corpus is subject to frequency analysis. even if a pre-processed text is almost uncomprehensible to a reader (specially if stopwords are removed), there probably is still a degree of trust in the analyst. she is usually someone who has no incentive to attempt a decipherment of the text or someone who has a lesser (but by no means inexisting) security clearance. this vulnerability will be investigated in the future.
• memory complexity is estimated to be at most double the size of the biggest document in the corpus.

@odanoburu & @fccoelho

LGPL 3, check the LICENSE.md file for full content.