[FEA] Word ngram based minhashes #15055
Assignees
Labels
feature request
New feature or request
libcudf
Affects libcudf (C++/CUDA) code.
strings
strings issues (C++ and Python)
Milestone
Comments
GregoryKimball
added
libcudf
3 tasks
rapids-bot bot
pushed a commit
that referenced
this issue
Sep 17, 2024
Experimental implementation for #15055 The input is a lists column of strings where each string in each row is expected as a word to be hashed. The minimum hash for that row is returned in a lists column where each row contains a minhash per input hash seed. Here the caller is expected to produce the words to be hashed. ``` std::unique_ptr<cudf::column> word_minhash( cudf::lists_column_view const& input, cudf::device_span<uint32_t const> seeds, rmm::cuda_stream_view stream, rmm::device_async_resource_ref mr); ``` Authors: - David Wendt (https://github.com/davidwendt) Approvers: - Bradley Dice (https://github.com/bdice) - Nghia Truong (https://github.com/ttnghia) - GALI PREM SAGAR (https://github.com/galipremsagar) URL: #15368
|
Word-based minhash API was deprecated in 24.12 and removed in 25.02. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Is your feature request related to a problem? Please describe.
The current minhashes functionality/API #12961 computes minhashes by taking a fixed character width sliding window and computes the minhashes for a document using character ngrams.
Another alternative approach to minhashing documents is to instead use word based ngrams instead of char based and research seems to suggest it leads to better quality results (lower false positives during Locality sensitive hashing).
Describe the solution you'd like
Add a word ngram based minhash support that's accelerated on GPUs.
Describe alternatives you've considered
Currently most CPU libraries achieve this by first tokenizing the string into word ngrams, and then looping over these tokens computing the minhash.
It is possible to somewhat mimic that with a
str.word_tokenize + str.hash_values + groupby + minin cuDF but requires a lot more intermediate memory and reduces the batch size of documents processed. It is also slower than a custom minhash implementation.Another challenge is that the definition of a "word" can be different for different languages and there's often different approaches to create word_ngrams based on language.
Additional context
Add any other context, code examples, or references to existing implementations about the feature request here.
The text was updated successfully, but these errors were encountered: