File size: 2,999 Bytes
f324ee0
 
 
 
 
 
da7a4c3
 
 
9dc288a
f324ee0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98fb129
f324ee0
 
 
 
 
 
 
 
e995672
f0204f8
e995672
f324ee0
 
 
 
 
 
 
 
 
 
 
 
 
e995672
f324ee0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
---
pretty_name: DPST Replication
language:
- en
tags:
- nlp
- data-privacy
- text-rewriting
- local-differential-privacy
viewer: false
---

# DPST Replication

Replication package for the EMNLP 2025 paper:

> **Leveraging Semantic Triples for Private Document Generation with Local Differential Privacy Guarantees**  
> Stephen Meisenbacher, Maulik Chevli, Florian Matthes  
> [https://aclanthology.org/2025.emnlp-main.455/](https://aclanthology.org/2025.emnlp-main.455/)

All credit goes to the original authors. This repository provides a replication environment for running the method, based on the official released code: https://github.com/sjmeis/DPST.

```bibtex
@inproceedings{meisenbacher-etal-2025-leveraging,
    title = "Leveraging Semantic Triples for Private Document Generation with Local Differential Privacy Guarantees",
    author = "Meisenbacher, Stephen and Chevli, Maulik and Matthes, Florian",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.455/",
    doi = "10.18653/v1/2025.emnlp-main.455",
    pages = "8976--8992",
}
```

## Setup

**1. Clone this repository**

```bash
git clone https://huggingface.co/datasets/weijunl/dpst-replication
cd dpst-replication
```

**2. Download and extract the Weaviate triple database**

```bash
tar -xzf weaviate-data.tar.gz && rm weaviate-data.tar.gz
```

**3. Create conda environment**

```bash
conda create -n dpst python=3.10
conda activate dpst

pip install numpy pandas tqdm torch accelerate
pip install transformers==4.52.4 datasets sentence-transformers==3.4.1
pip install nltk einops datasketch importlib_resources spacy
pip install stanford-openie
# install weaviate-client AFTER stanford-openie (protobuf conflict workaround)
pip install weaviate-client==4.11.1
```

**4. Start Weaviate (first time only — creates the container)**

```bash
docker run -d --name weaviate -p 8080:8080 -p 50051:50051 \
  -v ./weaviate-data:/var/lib/weaviate \
  -e AUTHENTICATION_ANONYMOUS_ACCESS_ENABLED=true \
  -e PERSISTENCE_DATA_PATH=/var/lib/weaviate \
  -e DEFAULT_VECTORIZER_MODULE=none \
  -e CLUSTER_HOSTNAME=node1 \
  cr.weaviate.io/semitechnologies/weaviate:1.26.4
```

After the first time, `run_dpst.sh` will automatically start the existing container if it is not running.

## Usage

```bash
conda activate dpst
bash run_dpst.sh <mode> <epsilon> [input_file] [output_file]

  mode     : 50k | 100k | 200k
  epsilon  : privacy budget (e.g. 0.1, 1.0, 10.0)
  input    : path to a text file (one sentence per line); omit for demo texts
  output   : path to save privatized output; omit to print to stdout
```

Example:
```bash
bash run_dpst.sh 200k 1.0 input.txt output.txt
```

A HuggingFace token is required for the generation model (Llama-3.2); `run_dpst.sh` will prompt for it at runtime.