Interoperability of Locality-Sensitive Hashing Schemes
This (work-in-progress) document proposes specific document fingerprinting algorithms, configuration, and practices for use in identifying and cross-referencing nearly identical documents in several medias. For example, web content across various HTML encodings and web design revisions, or research publications across PDF, HTML, and XML formats.
The motivation to come to a public consensus on specific details of these techniques is to enhance collaboration in efforts like web archiving, automated document identification, verifying repository contents, and near-duplicate-detection across distributed datasets.
These techniques are variously refered to as "near duplicate detection", "locality-sensitive hashing", and "similarity hashing". The major advances in this area historically came from commercial search engine development and the field of information retrieval in academia.
Contributions, corrections, and feedback to this document welcomed! You can use github features (issues and pull-requests) or email contributors directly.
The techniques described here usually complement both regular content hashing and full-document similarity measures. Regular hashing (or digests) generally have the property that exact byte-for-byte file matches will have equal hashes, but any small difference between files results in a completely different hash. The hash is of a fixed length of some 40 to 64 bytes (very small compared to document size). This is useful for fast equality checking and verify the integrity of files, but is too brittle for use identification across media. Similarity measures (such as cosine similarity or Hamming distance) often take a full document (or an extract set of tokens from the document) and give a fractional measure of similarity. They have the property of being relatively inefficient for lookups in large corpuses because each document must be compared one-to-one.
Similarity hashing is often used as an optimization on top of similarity
measures: the hashes of all documents are pre-computed and then sorted, and
only those documents within a fixed distance in the sorted list need to be
compared. This reduces identification of near-duplicates in a corpus from
O(N^2) similarity computations to
O(N) similarity compurations. Indexed
lists also reduce work in the context of individual lookup query for large
Some notable algorithms to consider are:
- simhash (aka, Charikar hashing)
- weighed minhash
- shingle methods (with dozens of feature hashes per document)
However, the papers describing these algorithms have a few parameters (eg, hash bit length, token hash algorithm), and leave several important steps up to the implementor, such as:
- media-specific content extraction (eg, stripping some or all HTML tags)
- input encoding normalization (UTF-8, etc)
- tokenization (particularly for non-latin languages)
- stop words
- string serialization (eg, base16, base32, or base64 encoding)
- prefixes or short identifiers for specific schemes
These parameters are usually left to be tuned by implementors. Some parameters can still be left to tuning by implementors, because they happen at comparison time, not during batch indexing:
- number of segmented rotations when sorting and distance of match (in bit flips) to seach for
- algorithm or measure used for final similarity measurement
Note that unlike regular hashing algorithms, the details for similarity hashes don't necessarily need to be specified in completeness: if differences between implementations only result in (statistically) tiny changes in the hash, hashes should still compare somewhat accurately. Complete reproducibility (exact similarity hash output for exact document input) between implementations is desirable, but not strictly necessary in all cases, leaving some wiggle room for implementors and optimization.
Similar to the
SHA families of hashes, I can imagine that a
small (1-4) set of variants might be useful for use in different contexts. It's
also good practice to build software, protocols, and data models to permit
swapping out algorithms for future flexibility, though of course the whole
concept here is to settle on a small number of consensus schemes for
TODO: summary table here
The roughly defined scope for this scheme would be "documents" of length between one and fifty pages when printed. If it could scale down to 2-3 paragraphs and up to thousand-page documents that would be great, but it should be optimized for the narrower scope. It should be useful across media (web pages, PDFs, plain text, etc), and across the most popular languages.
Proposed extraction and tokenization:
- strip all metadata and styling when extracting. include figure captions, table contents, quoted text. Do include reference lists, but do not include tokens from URLs or identifiers.
- UTF-8 encoded tokens
- fallback to unicode word-character boundaries for tokenization if a language-specific tokenizer is not available
- tokens should include only "word characters", as commonly included in
unicode-aware regex libraries. Specifically including the cateogires:
Ll Lu Lt Lo Lm Mn Nd Pc. They must include at least one letter/"Alphabetic" character.
- specifically, no zero-width or non-printing unicode modifiers
- numbers (unless part of an alphanumeric string, eg an acronym) should not be included
- TODO: instead, just strip all numeric characters?
- OPTIONALLY, a language-specific stop-list appropriate for search-engine indexing may be used.
Proposed algorithm and parameters:
jenkinshash algorithm for tokens
- 64x signed buckets (of 64-bit depth) during calculation
- 64-bit final form
- base32 when serialized as a string (non-capitalization-sensitive)
Note: endianness does not need to be specified, as the 64bit feature hashes are defined as 8-byte octet strings, not 64-bit integers.
Recommended default lookup parameters (these are all optional):
- k=3 ("distance", or number of differing bits to allow when filtering)
- Hamming distance as a similarity metric (for speed)
- 0.90 considered a loose match, 0.98 considered a close match
- create 16 tables of 28-bits for lookup, by first chosing 1 out of 4 16-bit blocks, then subdividing the remaining 48 bits into 4 12-bit blocks and choosing one of them, as described in Manku et al 3.1.1.
Existing Libraries and Implementations
These don't necessarily implement the schemes above, but they do implement the basic algorithms.
- seomoz/simhash-cpp (C++) and seomoz/simhash-py (python, wrapping the C++ library): simhash, used in production for web de-duplication
- simhash (C++): mirror of Google's simhash implementation (old?)
- datasketch (Python): minhash, Jaccard similarity, minhash-based indexing with Jaccard threshold, Top-K, or containment threshold.
- sean-public/python-hashes (Python)
- sing1ee/simhash-java (Java)
- codelibs/elasticsearch-minhash: plugin for a general-purpose search engine
- bbalet/stopwords (Golang): for a dozen+ languages. also does HTML stripping
"Near-Duplicate Detection", Lecoc. 2015. https://moz.com/devblog/near-duplicate-detection/
Moz blogpost. Good place to start.
"Charikar: Similarity Estimation Techniques from Rounding Algorithms", in Proceedings of the thiry-fourth annual ACM symposium on Theory of computing, ACM Press, 2002"
The original simhash paper
"Manku, Jain, Sarma: Detecting Near-Duplicates for Web Crawling". in Proceedings of the 16th international conference on World Wide Web, ACM Press, 2007
"Identifying and Filtering Near-Duplicate Documents", Broder. 2000.
"Near Duplicate Detection in an Academic Digital Library", Williams, Giles. 2013. http://www.personal.psu.edu/kiw5209/papers/2013/williams_doceng2013.pdf
"Probabilistic Near-Duplicate Detection Using Simhash", Sood and Loguinov. 2011.
"In Defense of MinHash over SimHash" http://proceedings.mlr.press/v33/shrivastava14.pdf
"Finding Similar Files in a Large File System", Manber. 1993. http://webglimpse.net/pubs/TR93-33.pdf
For a bibliography of older work, see https://github.com/JanX2/simhash/blob/master/paper/simHashBiblio.bib.