Identifying documents that contain potential code words using a machine learning model
Identifying documents that contain potential code words using a machine learning model. In some embodiments, a method may include receiving documents, identifying a first corpus and a second corpus in the documents, extracting a first set of word embeddings from the first corpus and a second set of word embeddings from the second corpus, generating a first vector space for the first set of word embeddings and a second vector space for the second set of word embeddings using a machine learning model, performing a vector rotation to improve alignment of the first set of word embeddings with the second set of word embeddings, identifying a word embedding in the first vector space that is not aligned with a corresponding word embedding in the second vector space as a potential code word, and identifying one or more documents that contain the potential code word in the first corpus.
1. A computer-implemented method for identifying documents that contain potential code words using a machine learning model, at least a portion of the method being performed by a computer system comprising one or more processors, the method comprising:
obtaining a plurality of document;
identifying, by the computer system, a plurality of document corpuses from the plurality of documents;
extracting, by the computer system, a set of word embeddings from each of the plurality of document corpuses such that a plurality of sets of word embeddings are extracted;
generating, by the computer system, a plurality of vector spaces for the sets of word embeddings using a machine learning model, with each set of word embeddings having a unique vector space;
grouping, by the computer system, the plurality of vector spaces in pairs, such that a first vector space of the plurality of vector spaces is grouped with a second vector space of the plurality of vector spaces in a first pair and a third vector space of the plurality of vector spaces in a second pair;
comparing, by the computer system, the word embeddings of the vector spaces that are paired;
identifying, by the computer system, a word embedding as a potential code phrase that includes one or more words based on the comparisons of the word embeddings of the vector spaces that are paired; and
identifying, by the computer system, one or more documents of the plurality of documents that contain the potential code phrase, the potential code phrase having a different meaning in the one or more documents than a standard meaning of the potential code phrase.
2. The method of claim 1 , wherein the comparing, by the computer system, the word embeddings of the vector spaces that are paired generates an alignment score for each pair of vector spaces and the word embedding is identified as the potential code phrase based on the alignment score between the vector spaces that are paired being below a threshold.
3. The method of claim 2 , further comprising:
in response to a comparison of the word embeddings of a pair of vector spaces having a comparison score above a threshold, combining the vector spaces of the pair of vector spaces;
grouping the combined vector spaces with another vector space of the plurality of vector spaces in a third pair; and
comparing the word embeddings of the vector spaces of the third pair,
wherein the word embedding is identified as the potential code phrase is further based on the comparison of the word embeddings of the vector spaces of the third pair.
4. The method of claim 1 , further comprising:
assigning a comparison score to each of the vector space pairs based on the comparisons of the word embeddings of the vector spaces that are paired; and
ranking the vector space pairs based on the comparison scores,
wherein the word embedding is identified as the potential code phrase based on the ranking of the vector space pairs.
5. The method of claim 1 , further comprising before grouping the plurality of vector spaces in pairs, selecting a first group of the plurality of vector spaces, wherein grouping the plurality of vector spaces in pairs includes grouping each of the plurality of vector spaces not included in the first group with at least one of the vector spaces in the first group to form vector space pairs.
6. The method of claim 1 , further comprising performing a vector rotation of one or more vector spaces of the plurality of vector spaces to improve alignment of the plurality of vector spaces.
7. The method of claim 1 , wherein the generating of the plurality of vector spaces using a machine learning model further comprises generating the plurality of vector spaces using a Word2Vec machine learning model.
8. The method of claim 1 , wherein comparing the word embeddings of the vector spaces that are paired comprises performing a cosine similarity comparison.
9. The method of claim 1 , wherein the documents include emails.
10. One or more non-transitory computer-readable media comprising one or more computer-readable instructions that, when executed by one or more processors of a system, cause the system to perform the method of claim 1 .
11. A system comprising:
one or more non-transitory computer-readable media comprising one or more computer-readable instructions; and
one or more processors configured to execute the computer-readable instructions to cause the system to perform operations, the operations comprising:
obtaining a plurality of document;
identifying a plurality of document corpuses from the plurality of documents;
extracting a set of word embeddings from each of the plurality of document corpuses such that a plurality sets of word embeddings are extracted;
generating a plurality of vector spaces for the sets of word embeddings using a machine learning model, with each set of word embeddings having a unique vector space;
grouping the plurality of vector spaces in pairs, such that a first vector space of the plurality of vector spaces is grouped with a second vector space of the plurality of vector spaces in a first pair and a third vector space of the plurality of vector spaces in a second pair;
comparing the word embeddings of the vector spaces that are paired;
identifying a word embedding as a potential code phrase that includes one or more words based on the comparisons of the word embeddings of the vector spaces that are paired; and
identifying one or more documents of the plurality of documents that contain the potential code phrase, the potential code phrase having a different meaning in the one or more documents than a standard meaning of the potential code phrase.
12. The system of claim 11 , wherein the comparing the word embeddings of the vector spaces that are paired generates an alignment score for each pair of vector spaces and the word embedding is identified as the potential code phrase based on the alignment score between the vector spaces that are paired being below a threshold.
13. The system of claim 12 , wherein the operations further comprise:
in response to a comparison of the word embeddings of a pair of vector spaces having a comparison score above a threshold, combining the vector spaces of the pair of vector spaces;
grouping the combined vector spaces with another vector space of the plurality of vector spaces in a third pair; and
comparing the word embeddings of the vector spaces of the third pair,
wherein the word embedding is identified as the potential code phrase is further based on the comparison of the word embeddings of the vector spaces of the third pair.
14. The system of claim 11 , wherein the operations further comprise:
assigning a comparison score to each of the vector space pairs based on the comparison of the word embeddings of the vector spaces that are paired; and
ranking the vector space pairs based on the comparison scores,
wherein the word embedding is identified as the potential code phrase based on the ranking of the vector space pairs.
15. The system of claim 11 , wherein the operations further comprise before grouping the plurality of vector spaces in pairs, selecting a first group of the plurality of vector spaces, wherein grouping the plurality of vector spaces in pairs includes grouping each of the plurality of vector spaces not included in the first group with at least one of the vector spaces in the first group to form vector space pairs.
16. The system of claim 11 , wherein the operations further comprise performing a vector rotation of one or more vector spaces of the plurality of vector spaces to improve alignment of the plurality of vector spaces.
17. The system of claim 11 , wherein the generating of the plurality of vector spaces using a machine learning model further comprises generating the plurality of vector spaces using a Word2Vec machine learning model.
18. The system of claim 11 , wherein comparing the word embeddings of the vector spaces that are paired comprises performing a cosine similarity comparison.
19. The system of claim 11 , wherein the plurality of documents include emails.