Skip to content

Latest commit

 

History

History
79 lines (62 loc) · 3.57 KB

Readme.md

File metadata and controls

79 lines (62 loc) · 3.57 KB
Raw

DReditor: An Time-efficient Approach for Building a Domain-specific Dense Retrieval Model

Co-author: Duanyu Feng (Sichuan University), Chen Huang (Sichuan University)

This is the code of our model DReditor, which is an computation-effective approach for domain-specific Dense Retrieval (DR) with training free.

Paper

Click Me

Requirements

Our framework is based on the previous work, BEIR, which is a heterogeneous benchmark containing diverse IR tasks. Therefore, the requirements of our framework are the same as the BEIR. See more detail in

beir-cellar/beir: A Heterogeneous Benchmark for Information Retrieval. Easy to use, evaluate your models across 15+ diverse IR datasets. (github.com)

DReditor

The core of our framework is an editor layer, which is a simple plug-in code. The code is in the file Mylayereditor.py. With this file, you can use BEIR to test any Dense Retrieval models supported by BEIR.

An example:

from beir import util, LoggingHandler
from beir.retrieval import models
from beir.datasets.data_loader import GenericDataLoader
from beir.retrieval.evaluation import EvaluateRetrieval
from beir.retrieval.search.dense import DenseRetrievalExactSearch as DRES
from beir.retrieval.search.sparse import SparseSearch
from Mylayereditor import Mylayereditor
from CustomModel import YourCustomDEModel
import random
import logging
import pathlib, os

#### Just some code to print debug information to stdout
logging.basicConfig(format='%(asctime)s - %(message)s',
                    datefmt='%Y-%m-%d %H:%M:%S',
                    level=logging.INFO,
                    handlers=[LoggingHandler()])
#### print debug information to stdout
os.environ["CUDA_VISIBLE_DEVICES"]='5'
#### Download scifact.zip dataset and unzip the dataset
dataset = "scifact"
url = "https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/{}.zip".format(dataset)
out_dir = os.path.join(pathlib.Path(__file__).parent.absolute(), "datasets")
data_path = util.download_and_unzip(url, out_dir)

#### the data we need to edit for the origin model
corpus_edit, queries_edit, qrels_edit = GenericDataLoader(data_folder=data_path).load(split="train")

#### the data we need to test
corpus, queries, qrels = GenericDataLoader(data_folder=data_path).load(split="test")

#### Load the dpr model and retrieve using cosine-similarity
model = models.SentenceBERT(("facebook-dpr-question_encoder-multiset-base", "facebook-dpr-ctx_encoder-multiset-base", " [SEP] "))

#### our plug-in for editing the model
model_run = Mylayereditor(model, batch_size=16, calibrate_type="q")
delta = model_run.layer_editor(corpus_t, queries_t, qrels_t, limits=500)

#### Evaluate the model with NDCG@k, MAP@K, Recall@K and Precision@K  where k = [1,3,5,10,100,1000]
model_run.turn_on()
retriever = EvaluateRetrieval(model_run, score_function="cos_sim") 
results = retriever.retrieve(corpus, queries)
ndcg, _map, recall, precision = retriever.evaluate(qrels, results, retriever.k_values)

Reference

If you make advantage of the DREditor in your research, please cite the following in your manuscript:

@misc{huang2024dreditor,
      title={DREditor: An Time-efficient Approach for Building a Domain-specific Dense Retrieval Model}, 
      author={Chen Huang and Duanyu Feng and Wenqiang Lei and Jiancheng Lv},
      year={2024},
      eprint={2401.12540},
      archivePrefix={arXiv},
      primaryClass={cs.IR}
}