Repository files navigation

iscc-sct is aSemantic-Code Text implementation for theISCC(International Standard Content Code). The Semantic-Code Text is a new ISCC-UNIT for semantic textidentification. The algorithm creates simmilar (low hamming distance) codes for semantically similartext inputs across different languages. The SCT ISCC-UNIT is a compact binary code created from abinarized document-vector text-embedding.

The ISCC is a combination of various similarity preserving fingerprints and an identifier fordigital media content.

ISCCs are generated algorithmically from digital content, just like cryptographic hashes. However,instead of using a single cryptographic hash function to identify data only, the ISCC uses variousalgorithms to create a composite identifier that exhibits similarity-preserving properties (softhash or Simprint).

The component-based structure of the ISCC identifies content at multiple levels of abstraction. Eachcomponent is self-describing, modular, and can be used separately or with others to aid in variouscontent identification tasks. The algorithmic design supports content deduplication, databasesynchronization, indexing, integrity verification, timestamping, versioning, data provenance,similarity clustering, anomaly detection, usage tracking, allocation of royalties, fact-checking andgeneral digital asset management use-cases.

The ISCC framework already includes a Text-Code based on lexical similarity for near-duplicatematching. The ISCC Semantic Text-Code is a planned additional ISCC-UNIT focused on capturing a moreabstract and broader semantic similarity. It is engineered to be robust against a wide range ofvariations and, most remarkably, translations of text that cannot be matched based on lexicalsimilarity alone.

One of the most interesting aspects of the Semantic Text-Code is its ability to generate(near)-identical codes for translations of the same text. This means that the same content,expressed in different languages, can be identified and linked, opening up new possibilities forcross-lingual content identification and similarity detection.

• Semantic Similarity: Utilizes deep learning models to generate codes that reflect the semanticessence of text.
• Translation Matching: Creates nearly identical codes for text translations, enablingcross-lingual content identification.
• Bit-Length Flexibility: Supports generating codes of various bit lengths (up to 256 bits),allowing for adjustable granularity in similarity detection.
• ISCC Compatible: Generates codes fully compatible with the ISCC specification, facilitatingseamless integration with existing ISCC-based systems.

Ensure you have Python 3.9 or newer installed on your system. Install the library using:

pip install iscc-sct

For systems with GPU CUDA support, enhance performance by installing with:

pip install iscc-sct[gpu]

Generate a Semantic Text-Code using the create function:

>>> import iscc_sct as sct>>> text = "This is some sample text. It can be a longer document or even an entire book.">>> sct.create(text, bits=256){  "iscc": "ISCC:CADV3GG6JH3XEVRNSVYGCLJ7AAV3BOT5J7EHEZKPFXEGRJ2CTWACGZI",  "characters": 77}

For granular (per chunk) feature outputs:

>>> import iscc_sct as sct>>> text = "This is some sample text. It can be a longer document or even an entire book.">>> sct.create(text, bits=256, granular=True){  "iscc": "ISCC:CADV3GG6JH3XEVRNSVYGCLJ7AAV3BOT5J7EHEZKPFXEGRJ2CTWACGZI",  "characters": 77,  "features": [    {      "maintype": "semantic",      "subtype": "text",      "version": 0,      "simprints": [        {          "simprint": "XZjeSfdyVi0",          "offset": 0,          "size": 77,          "content": "This is some sample text. It can be a longer document or even an entire book."        }      ]    }  ]}

The installation also provides a sct command-line tool:

usage: sct [-h] [-b BITS] [-g] [-d] [path]Generate Semantic Text-Codesfor text files.positional arguments:  path                  Path to text files (supports glob patterns) or'gui' to launch Gradio demo.options:  -h, --help            show thishelp message andexit  -b BITS, --bits BITS  Bit-Length of Code (default 256)  -g, --granular        Activate granular processing.  -d, --debug           Show debugging messages.

iscc-sct employs the following process:

1. Splits the text into overlaping chunks (using syntactically sensible breakpoints).
2. Uses a pre-trained deep learning model for text embedding.
3. Generates feature vectors capturing essential characteristics of the chunks.
4. Aggregates these vectors and binarizes them to produce a Semantic Text-Code.
5. Prefixes the binarized vector with the matching ISCC header, encodes it with base32, and adds the"ISCC:" prefix.

This process ensures robustness to variations and translations, enabling cross-lingual matchingbased on a short Simprint.

We welcome contributions to enhance the capabilities and efficiency of this proof of concept. Fordevelopment, install the project in development mode usingPoetry:

git clone https://github.com/iscc/iscc-sct.gitcd iscc-sctpoetry install

If you have suggestions for improvements or bug fixes, please open an issue or pull request. Formajor changes, please open an issue first to discuss your ideas.

We particularly welcome recommendations for other multilingual text embedding models trained withMatryoshka Representation Learning (MRL) and optimized for binarization. Such contributions couldsignificantly improve the performance and efficiency of the ISCC Semantic Text-Code generation.

This repository also provides an interactive Gradio demo that allows you to explore the capabilitiesof ISCC Semantic Text-Code. The demo showcases:

• Generation of ISCC Semantic Text-Codes for input texts
• Comparison of two texts and their similarity based on the generated codes
• Visualization of text chunking and granular matches
• Adjustable parameters like ISCC bit-length and maximum tokens per chunk

You can access the live version of the Gradio demo at:https://huggingface.co/spaces/iscc/iscc-sct

To run the Gradio demo locally, you first need to install theiscc-sct package with the optionaldemo dependency:

pip install iscc-sct[demo]

This will ensure that Gradio and other necessary dependencies for the demo are installed.

After installation, you can use thesct command-line tool that comes with the package:

sct gui

This command will launch the Gradio interface in your default web browser, allowing you to interactwith the demo on your local machine.

Arabic, Armenian, Bengali, Bosnian, Bulgarian, Burmese, Catalan, Chinese (China), Chinese (Taiwan),Croatian, Czech, Danish, Dutch, English, Estonian, Farsi, Finnish, French, French (Canada),Galician, German, Greek, Gujarati, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Italian,Japanese, Kannada, Korean, Kurdish, Latvian, Lithuanian, Macedonian, Malay, Malayalam, Marathi,Mongolian, Norwegian Bokmål, Persian, Polish, Portuguese, Portuguese (Brazil), Romanian, Russian,Serbian, Sinhala, Slovak, Slovenian, Spanish, Swedish, Tamil, Telugu, Thai, Turkish, Ukrainian,Urdu, Vietnamese.

The current chunking strategy uses tries to maximize chunk sizes (up to 127 tokens) wheil stillsplitting at lexically sensible boundaries with an overlap of up to 48 tokens. Seetext-splitter.

Cross document chunk matching via granular Simprints can likely be improved significantly with asemantically aware and shift resistant chunking strategy. Better shift resistance would improve thechances that the bounderies detected for semantically similar text sequences in different documentsare aligned.

A text embedding model trained withMatryoshka Representation Learning may yield better results withshort 64-bit Semantic Text-Codes.

A text embedding model with support for a largermax_token size (currently 128) may yieldhigher-order granular simprints based on larger chunks of text.

• Text Chunking:text-splitter
• Text Embeddings:Sentence-Transformers