Spaces:

---

TdelaSelle

/

Paraphrasis

like 0

Running on Zero

Dear Hugging Face team,

I am applying for a community GPU grant for Paraphrasis (TdelaSelle/Paraphrasis), an open-source (CC-BY-4.0) academic tool developed at the CNRS for researchers in the humanities and social sciences (HSS).

Paraphrasis measures semantic textual similarity (STS) between sequences from two corpora in order to detect textual reuse, intertextuality, paraphrase and quotation — a central task in philology, history and literary studies. It pairs a sentence-embedding back-end (Sentence-BERT bi-encoder + cosine similarity) with a manuscript-styled annotation interface so that scholars with no programming background can review and curate results. The default checkpoint (TdelaSelle/PatriSBERT) targets biblical reuse in Latin; the Space also accepts general multilingual sentence-transformer models.

Because dedicated sentence-embedding models barely exist for ancient and historical languages, Paraphrasis offers an optional unsupervised in-domain adaptation step: a short contrastive (SimCSE) fine-tuning run on sequences extracted from the user's own input files, which specialises the encoder to the vocabulary and register of the corpus before detection. This step, plus the encoding of both corpora, is the compute bottleneck. On the free cpu-basic tier it takes minutes and makes the interactive workflow impractical; on a GPU the same fine-tuning completes in well under a minute.

The workload is light and bursty: small input files, few concurrent users, GPU needed only for short fine-tuning + encoding bursts (peak VRAM well under 16 GB), idle the rest of the time. Even a modest GPU allocation, would let the research community use the tool as intended.
Remark: Paraphrasis is build on Docker and Streamlit, ZeroGPU solution on HF Spaces is excluded. As migration to the Gradio SDK is requiring significant efforts, we kindly ask for another GPU solution such as NVIDIA T4 small.

Thank you for considering this request, which directly serves the open-science goal of making AI methods accessible to non-technical HSS researchers.

— T. de la Selle, National Center for Scientific Research (CNRS)

Project overview

Paraphrasis addresses a long-standing problem in the text-based humanities: locating textual reuse (quotation, allusion, paraphrase, loci paralleli) across a source corpus and a reference corpus, through detection or information retrieval tasks. Traditional approaches rely on lexical overlap (n-gram or string matching), which fails precisely where reuse is most interesting — when wording is altered, abbreviated, translated or morphologically inflected. Reformulating the problem as semantic textual similarity (STS) rather than surface similarity is what makes neural sentence representations valuable here.

The tool is deliberately designed for users “who are not familiar with numerical aspects while expecting great insights from AI methods.” It exposes only several decisions — the two corpora, an embedding model, segmentation and tasks settings — and returns a ranked table of candidate pairs that scholars validate in an integrated annotation panel. The pipeline, input handling (.txt, .csv, .tsv), clause or semantic segmentation, encoding and review interface are already implemented and running on CPU; the grant concerns compute, not development.

Paraphrasis has already been presented in several conferences and a scientific community is already using it. A methodological publication presenting the approach based on bi-encoders and cosine similarity, the unsupervised sentence embedding learning and the experiments on text reuse detection is currently under review for the Journal of Data Mining and Digital Humanities (JDMDH). See arXiv version: https://arxiv.org/pdf/2607.24542

Community benefit and open-science alignment

Paraphrasis is released under CC-BY-4.0, with public code, a public Space, and example corpora in several languages on the Hub. It lowers the barrier between modern representation-learning methods and a research community (HSS, classics, biblical studies, history) that rarely has the engineering resources to apply them. This is squarely aligned with Hugging Face's mission to democratise AI: the grant would let an existing, working, open tool actually be usable by the scholars it is built for, rather than gated behind compute they do not have. The unsupervised, label-free adaptation method is itself a reusable recipe for other low-resource and historical-language projects on the Hub.