This repository contains a collection of text deduplication scripts that are ready to use, or modify based on your needs:
- MinHash + MinHashLSH, including a spark implementation suitable for large (TB) datasets
- 64 or 128 bit SimHash
- SuffixArray Substring
- Bloom Filter
- Exact Hash (document-level, line-level/ccnet)
I also have big plans for the future:
- Memory benchmark for streaming processing
- Inter-dataset deduplication
- Rewrite suffix array in Python
- A collections of other deduplication methods: SuperMinHash, ProbMinHash, TreeMinHash, BagMinHash, Optimal Densification for Fast and Accurate Minwise Hashing, Fast Similarity Sketching
However, I do not intent to build a general purpose deduplication library, which was the goal of this repo early on. I will gradually retire the pypi package as well. The reason behind it is that each use-case can be wildly different and requires careful design and consideration. I sincerely encourage you to read the script first (they are relatively short) so you can understand what are at stake here when using it. You can use it to bootstrap your own script, or just use it as a reference.
This repository is inspired by the following projects, and is heavily influenced by lessons learned from my own participation in BigScience (Apache 2.0) and BigCode (Apache 2.0). There is a blog post about the journey. Feedbacks are welcome!
- Datasketch (MIT)
- simhash-py and simhash-cpp (MIT)
- Deduplicating Training Data Makes Language Models Better (Apache 2.0)
- Gaoya (MIT)
Not a lot of people have access to enough compute resources or the need to deduplicate TB-scale datasets, but if you do, this is a good example of how to use it with GCP DataProc.
MODIFY text_dedup/minhash_spark.py
FOR YOUR OWN PROJECT AND DATASET FIRST!
export CLUSTER_NAME=chenghao-temp
export PROJECT_ID=xx
export REGION=us-central1
export ZONE=us-central1-a
export INPUT_GCS_PATH="gs://chenghao-temp-exp/data/ada"
export OUTPUT_GCS_PATH="gs://chenghao-temp-exp/output/ada"
gcloud dataproc clusters create $CLUSTER_NAME \
--enable-component-gateway \
--region $REGION \
--zone $ZONE \
--master-machine-type c2d-standard-16 \
--master-boot-disk-size 500 \
--num-workers 10 \
--worker-machine-type c2d-standard-16 \
--worker-boot-disk-size 500 \
--image-version 2.0-debian10 \
--project $PROJECT_ID
gcloud dataproc jobs submit pyspark --cluster ${CLUSTER_NAME}\
--region $REGION \
--jars gs://spark-lib/bigquery/spark-3.3-bigquery-0.32.2.jar \
--driver-log-levels root=FATAL,__main__=DEBUG \
--properties="spark.executor.memory"="50g","spark.driver.memory"="8g","spark.executor.cores"="14" \
minhash_spark.py -- --input $INPUT_GCS_PATH --output $OUTPUT_GCS_PATH
For reference, the script finished deduplicating 42 million rows in less than 40 minutes with above settings (160 cores, 640GB memory in total), while the python version would take around 10 hours with a 80-core machine with 1.8TB memory.
In the following part, we are going to deduplicate one dataset: gl
subset of oscar-corpus/OSCAR-2201
.
# input
python -m text_dedup.suffix_array \
--path "oscar-corpus/OSCAR-2201" \
--name "gl" \
--split "train" \
--cache_dir "./cache" \
--output "output/suffix_array/oscar_gl_dedup" \
--column "text" \
--google_repo_path "/Users/chenghao/Downloads/Projects/text-dedup/deduplicate-text-datasets"
# output
INFO Loading : 2.75 seconds
INFO Preprocessing : 4.78 seconds
INFO SuffixArray : 98.29 seconds
INFO SelfSimilar : 4.24 seconds
INFO Restore : 0.25 seconds
INFO Deduplicate : 6.23 seconds
INFO Saving : 8.91 seconds
INFO Total : 125.45 seconds
INFO Before : 180332342 bytes (88803)
INFO After : 97646271 bytes (40404)
# input
python -m text_dedup.minhash \
--path "oscar-corpus/OSCAR-2201" \
--name "gl" \
--split "train" \
--cache_dir "./cache" \
--output "output/minhash/oscar_gl_dedup" \
--column "text" \
--batch_size 10000
# output
INFO Loading : 2.62 seconds
INFO MinHashing : 0.08 seconds
INFO Clustering : 2.20 seconds
INFO Filtering : 0.53 seconds
INFO Saving : 9.86 seconds
INFO Total : 15.29 seconds
INFO Data Number (before) : 88803
INFO Data Number (after) : 44124 (49.69%)
INFO Duplicate Number : 44679 (50.31%)
INFO π€ Happy Deduplicating π€
# input
python -m text_dedup.simhash \
--path "oscar-corpus/OSCAR-2201" \
--name "gl" \
--split "train" \
--cache_dir "./cache" \
--output "output/simhash/oscar_gl_dedup" \
--column "text" \
--batch_size 10000
# output
INFO Loading : 2.60 seconds
INFO SimHashing : 0.04 seconds
INFO Indexing : 28.88 seconds
INFO Filtering : 0.88 seconds
INFO Saving : 10.41 seconds
INFO Total : 42.80 seconds
INFO Data Number (before) : 88803
INFO Data Number (after) : 46163 (51.98%)
INFO Duplicate Number : 42640 (48.02%)
INFO π€ Happy Deduplicating π€
# input
python -m text_dedup.exact_hash \
--path "oscar-corpus/OSCAR-2201" \
--name "gl" \
--split "train" \
--cache_dir "./cache" \
--output "output/exact_hash/oscar_gl_dedup" \
--column "text" \
--batch_size 1000
# output
INFO Loading : 2.95s
INFO Processing : 3.79s
INFO Filtering : 0.10s
INFO Saving : 2.89s
INFO Total : 9.72s
INFO Before : 88803
INFO After : 47049
# input
python -m text_dedup.bloom_filter \
--path "oscar-corpus/OSCAR-2201" \
--name "gl" \
--split "train" \
--cache_dir "./cache" \
--output "output/bloom_filter/oscar_gl_dedup" \
--error_rate 1e-5 \
--column "text" \
--batch_size 1000
# output
INFO Loading : 2.72s
INFO Processing : 4.84s
INFO Filtering : 0.10s
INFO Saving : 2.88s
INFO Total : 10.54s
INFO Before : 88803
INFO After : 47045
A benchmark of different methods here can be found in benchmarks/wiki40.ipynb
. A notebook in evaluating MinHash on pinecone/core-2020-05-10-deduplication
can be found in benchmarks/pinecone.ipynb
.
For quick reference, here are the results:
Method | Precision | Recall | F1 | Time |
---|---|---|---|---|
MinHash | 0.9464 | 0.9446 | 0.9455 | 24s |
SimHash* | 0.9011 | 0.6959 | 0.7853 | 210s |
SimHash(Gyawali et al., LREC 2020) | 0.697 | 0.247 | 0.3647 | - |
Exact Title (my implementation) | 0.8302 | 0.5521 | 0.6632 | - |
Exact Title(Gyawali et al., LREC 2020) | 0.830 | 0.50 | 0.624 | - |
*Best SimHash result from benchmarks/hyperparameter.ipynb
.