--- pipeline_tag: sentence-similarity datasets: - ms_marco - sentence-transformers/msmarco-hard-negatives metrics: - recall tags: - passage-retrieval library_name: sentence-transformers base_model: facebook/xmod-base inference: false language: - multilingual - af - am - ar - az - be - bg - bn - ca - cs - cy - da - de - el - en - eo - es - et - eu - fa - fi - fr - ga - gl - gu - ha - he - hi - hr - hu - hy - id - is - it - ja - ka - kk - km - kn - ko - ku - ky - la - lo - lt - lv - mk - ml - mn - mr - ms - my - ne - nl - no - or - pa - pl - ps - pt - ro - ru - sa - si - sk - sl - so - sq - sr - sv - sw - ta - te - th - tl - tr - uk - ur - uz - vi - zh ---

DPR-XM

đŸ› ïž Usage | 📊 Evaluation | đŸ€– Training | 🔗 Citation |

đŸ’» Code | 📄 Paper

This is a **multilingual** dense single-vector bi-encoder model. It maps questions and paragraphs 768-dimensional dense vectors and can be used for semantic search. The model uses an [XMOD](https://huggingface.co/facebook/xmod-base) backbone, which allows it to learn from monolingual fine-tuning in a high-resource language, like English, and perform zero-shot retrieval across multiple languages. ## Usage Here are some examples for using DPR-XM with [Sentence-Transformers](#using-sentence-transformers), [FlagEmbedding](#using-flagembedding), or [Huggingface Transformers](#using-huggingface-transformers). #### Using Sentence-Transformers Start by installing the [library](https://www.SBERT.net): `pip install -U sentence-transformers`. Then, you can use the model like this: ```python from sentence_transformers import SentenceTransformer queries = ["Ceci est un exemple de requĂȘte.", "Voici un second exemple."] passages = ["Ceci est un exemple de passage.", "Et voilĂ  un deuxiĂšme exemple."] language_code = "fr_FR" #Find all codes here: https://huggingface.co/facebook/xmod-base#languages model = SentenceTransformer('antoinelouis/dpr-xm') model[0].auto_model.set_default_language(language_code) #Activate the language-specific adapters q_embeddings = model.encode(queries, normalize_embeddings=True) p_embeddings = model.encode(passages, normalize_embeddings=True) similarity = q_embeddings @ p_embeddings.T print(similarity) ``` #### Using FlagEmbedding Start by installing the [library](https://github.com/FlagOpen/FlagEmbedding/): `pip install -U FlagEmbedding`. Then, you can use the model like this: ```python from FlagEmbedding import FlagModel queries = ["Ceci est un exemple de requĂȘte.", "Voici un second exemple."] passages = ["Ceci est un exemple de passage.", "Et voilĂ  un deuxiĂšme exemple."] language_code = "fr_FR" #Find all codes here: https://huggingface.co/facebook/xmod-base#languages model = FlagModel('antoinelouis/dpr-xm') model.model.set_default_language(language_code) #Activate the language-specific adapters q_embeddings = model.encode(queries, normalize_embeddings=True) p_embeddings = model.encode(passages, normalize_embeddings=True) similarity = q_embeddings @ p_embeddings.T print(similarity) ``` #### Using Transformers Start by installing the [library](https://huggingface.co/docs/transformers): `pip install -U transformers`. Then, you can use the model like this: ```python from transformers import AutoTokenizer, AutoModel from torch.nn.functional import normalize def mean_pooling(model_output, attention_mask): """ Perform mean pooling on-top of the contextualized word embeddings, while ignoring mask tokens in the mean computation.""" token_embeddings = model_output[0] #First element of model_output contains all token embeddings input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float() return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9) queries = ["Ceci est un exemple de requĂȘte.", "Voici un second exemple."] passages = ["Ceci est un exemple de passage.", "Et voilĂ  un deuxiĂšme exemple."] language_code = "fr_FR" #Find all codes here: https://huggingface.co/facebook/xmod-base#languages tokenizer = AutoTokenizer.from_pretrained('antoinelouis/dpr-xm') model = AutoModel.from_pretrained('antoinelouis/dpr-xm') model.set_default_language(language_code) #Activate the language-specific adapters q_input = tokenizer(queries, padding=True, truncation=True, return_tensors='pt') p_input = tokenizer(passages, padding=True, truncation=True, return_tensors='pt') with torch.no_grad(): q_output = model(**encoded_queries) p_output = model(**encoded_passages) q_embeddings = mean_pooling(q_output, q_input['attention_mask']) q_embedddings = normalize(q_embeddings, p=2, dim=1) p_embeddings = mean_pooling(p_output, p_input['attention_mask']) p_embedddings = normalize(p_embeddings, p=2, dim=1) similarity = q_embeddings @ p_embeddings.T print(similarity) ``` *** ## Evaluation - **mMARCO**: We evaluate our model on the small development sets of [mMARCO](https://huggingface.co/datasets/unicamp-dl/mmarco), which consists of 6,980 queries for a corpus of 8.8M candidate passages in 14 languages. Below, we compared its multilingual performance with other retrieval models on the dataset official metrics, i.e., mean reciprocal rank at cut-off 10 (MRR@10). | | model | Type | #Samples | #Params | en | es | fr | it | pt | id | de | ru | zh | ja | nl | vi | hi | ar | Avg. | |---:|:----------------------------------------------------------------------------------------------------------------------------------------|:--------------|:--------:|:-------:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:|-----:| | 1 | BM25 ([Pyserini](https://github.com/castorini/pyserini)) | lexical | - | - | 18.4 | 15.8 | 15.5 | 15.3 | 15.2 | 14.9 | 13.6 | 12.4 | 11.6 | 14.1 | 14.0 | 13.6 | 13.4 | 11.1 | 14.2 | | 2 | mono-mT5 ([Bonfacio et al., 2021](https://doi.org/10.48550/arXiv.2108.13897)) | cross-encoder | 12.8M | 390M | 36.6 | 31.4 | 30.2 | 30.3 | 30.2 | 29.8 | 28.9 | 26.3 | 24.9 | 26.7 | 29.2 | 25.6 | 26.6 | 23.5 | 28.6 | | 3 | mono-mMiniLM ([Bonfacio et al., 2021](https://doi.org/10.48550/arXiv.2108.13897)) | cross-encoder | 80.0M | 107M | 36.6 | 30.9 | 29.6 | 29.1 | 28.9 | 29.3 | 27.8 | 25.1 | 24.9 | 26.3 | 27.6 | 24.7 | 26.2 | 21.9 | 27.8 | | 4 | [DPR-X](https://huggingface.co/eugene-yang/dpr-xlmr-large-mtt-neuclir) ([Yang et al., 2022](https://doi.org/10.48550/arXiv.2204.11989)) | single-vector | 25.6M | 550M | 24.5 | 19.6 | 18.9 | 18.3 | 19.0 | 16.9 | 18.2 | 17.7 | 14.8 | 15.4 | 18.5 | 15.1 | 15.4 | 12.9 | 17.5 | | 5 | [mE5-base](https://huggingface.co/intfloat/multilingual-e5-base) ([Wang et al., 2024](https://doi.org/10.48550/arXiv.2402.05672)) | single-vector | 5.1B | 278M | 35.0 | 28.9 | 30.3 | 28.0 | 27.5 | 26.1 | 27.1 | 24.5 | 22.9 | 25.0 | 27.3 | 23.9 | 24.2 | 20.5 | 26.5 | | 6 | mColBERT ([Bonfacio et al., 2021](https://doi.org/10.48550/arXiv.2108.13897)) | multi-vector | 25.6M | 180M | 35.2 | 30.1 | 28.9 | 29.2 | 29.2 | 27.5 | 28.1 | 25.0 | 24.6 | 23.6 | 27.3 | 18.0 | 23.2 | 20.9 | 26.5 | | | | | | | | | | | | | | | | | | | | | | | 7 | **DPR-XM** (ours) | single-vector | 25.6M | 277M | 32.7 | 23.6 | 23.5 | 22.3 | 22.7 | 22.0 | 22.1 | 19.9 | 18.1 | 18.7 | 22.9 | 18.0 | 16.0 | 15.1 | 21.3 | | 8 | [ColBERT-XM](https://huggingface.co/antoinelouis/colbert-xm) (ours) | multi-vector | 6.4M | 277M | 37.2 | 28.5 | 26.9 | 26.5 | 27.6 | 26.3 | 27.0 | 25.1 | 24.6 | 24.1 | 27.5 | 22.6 | 23.8 | 19.5 | 26.2 | *** ## Training #### Data We use the English training samples from the [MS MARCO passage ranking](https://ir-datasets.com/msmarco-passage.html#msmarco-passage/train) dataset, which contains 8.8M passages and 539K training queries. We do not employ the BM25 netaives provided by the official dataset but instead sample harder negatives mined from 12 distinct dense retrievers, using the [msmarco-hard-negatives](https://huggingface.co/datasets/sentence-transformers/msmarco-hard-negatives) distillation dataset. Our final training set consists of 25.6M (q, p+, p-) triples. #### Implementation The model is initialized from the [xmod-base](https://huggingface.co/facebook/xmod-base) checkpoint and optimized via the in-batch sampled softmax cross-entropy loss (as in [DPR](https://doi.org/10.48550/arXiv.2004.04906)). It is fine-tuned on one 32GB NVIDIA V100 GPU for 200k steps using the AdamW optimizer with a batch size of 128, a peak learning rate of 2e-5 with warm up along the first 10\% of training steps and linear scheduling. We set the maximum sequence lengths for both the questions and passages to 128 tokens. *** ## Citation ```bibtex @article{louis2024modular, author = {Louis, Antoine and Saxena, Vageesh and van Dijck, Gijs and Spanakis, Gerasimos}, title = {ColBERT-XM: A Modular Multi-Vector Representation Model for Zero-Shot Multilingual Information Retrieval}, journal = {CoRR}, volume = {abs/2402.15059}, year = {2024}, url = {https://arxiv.org/abs/2402.15059}, doi = {10.48550/arXiv.2402.15059}, eprinttype = {arXiv}, eprint = {2402.15059}, } ```