opensearch-neural-sparse-encoding-doc-v1

Select the model

The model should be selected considering search relevance, model inference and retrieval efficiency(FLOPS). We benchmark models' zero-shot performance on a subset of BEIR benchmark: TrecCovid,NFCorpus,NQ,HotpotQA,FiQA,ArguAna,Touche,DBPedia,SCIDOCS,FEVER,Climate FEVER,SciFact,Quora.

Overall, the v2 series of models have better search relevance, efficiency and inference speed than the v1 series. The specific advantages and disadvantages may vary across different datasets.

Model	Inference-free for Retrieval	Model Parameters	AVG NDCG@10	AVG FLOPS
opensearch-neural-sparse-encoding-v1		133M	0.524	11.4
opensearch-neural-sparse-encoding-v2-distill		67M	0.528	8.3
opensearch-neural-sparse-encoding-doc-v1	✔️	133M	0.490	2.3
opensearch-neural-sparse-encoding-doc-v2-distill	✔️	67M	0.504	1.8
opensearch-neural-sparse-encoding-doc-v2-mini	✔️	23M	0.497	1.7

Overview

Paper: Towards Competitive Search Relevance For Inference-Free Learned Sparse Retrievers
Fine-tuning sample: opensearch-sparse-model-tuning-sample

This is a learned sparse retrieval model. It encodes the documents to 30522 dimensional sparse vectors. For queries, it just use a tokenizer and a weight look-up table to generate sparse vectors. The non-zero dimension index means the corresponding token in the vocabulary, and the weight means the importance of the token. And the similarity score is the inner product of query/document sparse vectors. In the real-world use case, the search performance of opensearch-neural-sparse-encoding-v1 is comparable to BM25.

This model is trained on MS MARCO dataset.

OpenSearch neural sparse feature supports learned sparse retrieval with lucene inverted index. Link: https://opensearch.org/docs/latest/query-dsl/specialized/neural-sparse/. The indexing and search can be performed with OpenSearch high-level API.

Usage (HuggingFace)

This model is supposed to run inside OpenSearch cluster. But you can also use it outside the cluster, with HuggingFace models API.

import json
import itertools
import torch

from transformers import AutoModelForMaskedLM, AutoTokenizer


# get sparse vector from dense vectors with shape batch_size * seq_len * vocab_size
def get_sparse_vector(feature, output):
    values, _ = torch.max(output*feature["attention_mask"].unsqueeze(-1), dim=1)
    values = torch.log(1 + torch.relu(values))
    values[:,special_token_ids] = 0
    return values
    
# transform the sparse vector to a dict of (token, weight)
def transform_sparse_vector_to_dict(sparse_vector):
    sample_indices,token_indices=torch.nonzero(sparse_vector,as_tuple=True)
    non_zero_values = sparse_vector[(sample_indices,token_indices)].tolist()
    number_of_tokens_for_each_sample = torch.bincount(sample_indices).cpu().tolist()
    tokens = [transform_sparse_vector_to_dict.id_to_token[_id] for _id in token_indices.tolist()]

    output = []
    end_idxs = list(itertools.accumulate([0]+number_of_tokens_for_each_sample))
    for i in range(len(end_idxs)-1):
        token_strings = tokens[end_idxs[i]:end_idxs[i+1]]
        weights = non_zero_values[end_idxs[i]:end_idxs[i+1]]
        output.append(dict(zip(token_strings, weights)))
    return output
    
# download the idf file from model hub. idf is used to give weights for query tokens
def get_tokenizer_idf(tokenizer):
    from huggingface_hub import hf_hub_download
    local_cached_path = hf_hub_download(repo_id="opensearch-project/opensearch-neural-sparse-encoding-doc-v1", filename="idf.json")
    with open(local_cached_path) as f:
        idf = json.load(f)
    idf_vector = [0]*tokenizer.vocab_size
    for token,weight in idf.items():
        _id = tokenizer._convert_token_to_id_with_added_voc(token)
        idf_vector[_id]=weight
    return torch.tensor(idf_vector)

# load the model
model = AutoModelForMaskedLM.from_pretrained("opensearch-project/opensearch-neural-sparse-encoding-doc-v1")
tokenizer = AutoTokenizer.from_pretrained("opensearch-project/opensearch-neural-sparse-encoding-doc-v1")
idf = get_tokenizer_idf(tokenizer)

# set the special tokens and id_to_token transform for post-process
special_token_ids = [tokenizer.vocab[token] for token in tokenizer.special_tokens_map.values()]
get_sparse_vector.special_token_ids = special_token_ids
id_to_token = ["" for i in range(tokenizer.vocab_size)]
for token, _id in tokenizer.vocab.items():
    id_to_token[_id] = token
transform_sparse_vector_to_dict.id_to_token = id_to_token



query = "What's the weather in ny now?"
document = "Currently New York is rainy."

# encode the query
feature_query = tokenizer([query], padding=True, truncation=True, return_tensors='pt', return_token_type_ids=False)
input_ids = feature_query["input_ids"]
batch_size = input_ids.shape[0]
query_vector = torch.zeros(batch_size, tokenizer.vocab_size)
query_vector[torch.arange(batch_size).unsqueeze(-1), input_ids] = 1
query_sparse_vector = query_vector*idf

# encode the document
feature_document = tokenizer([document], padding=True, truncation=True, return_tensors='pt', return_token_type_ids=False)
output = model(**feature_document)[0]
document_sparse_vector = get_sparse_vector(feature_document, output)


# get similarity score
sim_score = torch.matmul(query_sparse_vector[0],document_sparse_vector[0])
print(sim_score)   # tensor(12.8465, grad_fn=<DotBackward0>)


query_token_weight = transform_sparse_vector_to_dict(query_sparse_vector)[0]
document_query_token_weight = transform_sparse_vector_to_dict(document_sparse_vector)[0]
for token in sorted(query_token_weight, key=lambda x:query_token_weight[x], reverse=True):
    if token in document_query_token_weight:
        print("score in query: %.4f, score in document: %.4f, token: %s"%(query_token_weight[token],document_query_token_weight[token],token))
        

        
# result:
# score in query: 5.7729, score in document: 1.0552, token: ny
# score in query: 4.5684, score in document: 1.1697, token: weather
# score in query: 3.5895, score in document: 0.3932, token: now

The above code sample shows an example of neural sparse search. Although there is no overlap token in original query and document, but this model performs a good match.

Detailed Search Relevance

Model	Average	Trec Covid	NFCorpus	NQ	HotpotQA	FiQA	ArguAna	Touche	DBPedia	SCIDOCS	FEVER	Climate FEVER	SciFact	Quora
opensearch-neural-sparse-encoding-v1	0.524	0.771	0.360	0.553	0.697	0.376	0.508	0.278	0.447	0.164	0.821	0.263	0.723	0.856
opensearch-neural-sparse-encoding-v2-distill	0.528	0.775	0.347	0.561	0.685	0.374	0.551	0.278	0.435	0.173	0.849	0.249	0.722	0.863
opensearch-neural-sparse-encoding-doc-v1	0.490	0.707	0.352	0.521	0.677	0.344	0.461	0.294	0.412	0.154	0.743	0.202	0.716	0.788
opensearch-neural-sparse-encoding-doc-v2-distill	0.504	0.690	0.343	0.528	0.675	0.357	0.496	0.287	0.418	0.166	0.818	0.224	0.715	0.841
opensearch-neural-sparse-encoding-doc-v2-mini	0.497	0.709	0.336	0.510	0.666	0.338	0.480	0.285	0.407	0.164	0.812	0.216	0.699	0.837

License

This project is licensed under the Apache v2.0 License.