tomaarsen/splade-cocondenser-gooaq

splade-cocondenser trained on GooAQ

This is a SPLADE Sparse Encoder model finetuned from Luyu/co-condenser-marco on the gooaq dataset using the sentence-transformers library. It maps sentences & paragraphs to a 30522-dimensional sparse vector space and can be used for semantic search and sparse retrieval.

Model Details

Model Description

• Model Type: SPLADE Sparse Encoder
• Base model: Luyu/co-condenser-marco
• Maximum Sequence Length: 256 tokens
• Output Dimensionality: 30522 dimensions
• Similarity Function: Dot Product
• Training Dataset:
  • gooaq
• Language: en
• License: apache-2.0

Model Sources

• Documentation: Sentence Transformers Documentation
• Documentation: Sparse Encoder Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Sparse Encoders on Hugging Face

Full Model Architecture

SparseEncoder(
  (0): MLMTransformer({'max_seq_length': 256, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
  (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': 30522})
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SparseEncoder

# Download from the 🤗 Hub
model = SparseEncoder("tomaarsen/splade-cocondenser-gooaq")
# Run inference
sentences = [
    'what is the difference between 18 and 20 inch tires?',
    'The only real difference is a 20" rim would be more likely to be damaged, as you pointed out. Beyond looks, there is zero benefit for the 20" rim. Also, just the availability of tires will likely be much more limited for the larger rim. ... Tire selection is better for 18" wheels than 20" wheels.',
    'So extracurricular activities are just activities that you do outside of class. The Common App says that extracurricular activities "include arts, athletics, clubs, employment, personal commitments, and other pursuits."',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# (3, 30522)

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Evaluation

Metrics

Sparse Information Retrieval

• Datasets: NanoClimateFEVER, NanoDBPedia, NanoFEVER, NanoFiQA2018, NanoHotpotQA, NanoMSMARCO, NanoNFCorpus, NanoNQ, NanoQuoraRetrieval, NanoSCIDOCS, NanoArguAna, NanoSciFact and NanoTouche2020
• Evaluated with SparseInformationRetrievalEvaluator

Metric	NanoClimateFEVER	NanoDBPedia	NanoFEVER	NanoFiQA2018	NanoHotpotQA	NanoMSMARCO	NanoNFCorpus	NanoNQ	NanoQuoraRetrieval	NanoSCIDOCS	NanoArguAna	NanoSciFact	NanoTouche2020
dot_accuracy@1	0.18	0.6	0.74	0.34	0.72	0.24	0.42	0.44	0.8	0.36	0.14	0.48	0.6327
dot_accuracy@3	0.38	0.78	0.88	0.46	0.86	0.52	0.6	0.6	0.94	0.62	0.5	0.68	0.7959
dot_accuracy@5	0.52	0.84	0.94	0.52	0.9	0.62	0.66	0.68	0.98	0.7	0.6	0.7	0.8367
dot_accuracy@10	0.62	0.92	0.94	0.62	0.94	0.74	0.76	0.82	1.0	0.76	0.74	0.76	0.9592
dot_precision@1	0.18	0.6	0.74	0.34	0.72	0.24	0.42	0.44	0.8	0.36	0.14	0.48	0.6327
dot_precision@3	0.14	0.52	0.2933	0.2067	0.4267	0.1733	0.3467	0.2	0.38	0.2667	0.1667	0.2333	0.5374
dot_precision@5	0.12	0.512	0.2	0.152	0.28	0.124	0.328	0.144	0.248	0.224	0.12	0.148	0.502
dot_precision@10	0.08	0.452	0.102	0.096	0.15	0.074	0.282	0.09	0.134	0.166	0.074	0.086	0.4327
dot_recall@1	0.115	0.0552	0.7167	0.1856	0.36	0.24	0.0467	0.42	0.6973	0.074	0.14	0.455	0.0437
dot_recall@3	0.1983	0.1102	0.8267	0.3138	0.64	0.52	0.0979	0.56	0.8947	0.1667	0.5	0.65	0.1111
dot_recall@5	0.2683	0.1586	0.9167	0.3521	0.7	0.62	0.12	0.65	0.946	0.2317	0.6	0.68	0.1725
dot_recall@10	0.3323	0.2872	0.9233	0.4247	0.75	0.74	0.1616	0.79	0.99	0.3427	0.74	0.76	0.2839
dot_ndcg@10	0.2594	0.5254	0.8284	0.3557	0.6951	0.4858	0.3519	0.5968	0.8923	0.3212	0.439	0.6144	0.4877
dot_mrr@10	0.3062	0.6989	0.8097	0.4182	0.7948	0.4052	0.5292	0.545	0.8785	0.5074	0.3432	0.5732	0.7339
dot_map@100	0.2072	0.394	0.7934	0.306	0.6244	0.4218	0.168	0.5361	0.8493	0.2368	0.3522	0.5697	0.359

Sparse Nano BEIR

• Dataset: NanoBEIR_mean
• Evaluated with SparseNanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "climatefever",
          "dbpedia",
          "fever",
          "fiqa2018",
          "hotpotqa",
          "msmarco",
          "nfcorpus",
          "nq",
          "quoraretrieval",
          "scidocs",
          "arguana",
          "scifact",
          "touche2020"
      ]
  }
  

Metric	Value
dot_accuracy@1	0.4687
dot_accuracy@3	0.6628
dot_accuracy@5	0.7305
dot_accuracy@10	0.8138
dot_precision@1	0.4687
dot_precision@3	0.2993
dot_precision@5	0.2386
dot_precision@10	0.1707
dot_recall@1	0.273
dot_recall@3	0.4299
dot_recall@5	0.4935
dot_recall@10	0.5789
dot_ndcg@10	0.5272
dot_mrr@10	0.5803
dot_map@100	0.4475

Training Details

Training Dataset

gooaq

• Dataset: gooaq at b089f72
• Size: 3,011,496 training samples
• Columns: question and answer
• Approximate statistics based on the first 1000 samples:

  	question	answer
  type	string	string
  details	min: 8 tokens mean: 11.87 tokens max: 23 tokens	min: 14 tokens mean: 60.09 tokens max: 201 tokens

• Samples:

  question	answer
  what is the difference between clay and mud mask?	The main difference between the two is that mud is a skin-healing agent, while clay is a cosmetic, drying agent. Clay masks are most useful for someone who has oily skin and is prone to breakouts of acne and blemishes.
  myki how much on card?	A full fare myki card costs $6 and a concession, seniors or child myki costs $3. For more information about how to use your myki, visit ptv.vic.gov.au or call 1800 800 007.
  how to find out if someone blocked your phone number on iphone?	If you get a notification like "Message Not Delivered" or you get no notification at all, that's a sign of a potential block. Next, you could try calling the person. If the call goes right to voicemail or rings once (or a half ring) then goes to voicemail, that's further evidence you may have been blocked.

• Loss: SpladeLoss with these parameters:

  {'loss': SparseMultipleNegativesRankingLoss(
    (model): SparseEncoder(
      (0): MLMTransformer({'max_seq_length': 256, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
      (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': None})
    )
    (cross_entropy_loss): CrossEntropyLoss()
  ), 'lambda_corpus': 3e-05, 'lambda_query': 5e-05, 'corpus_regularizer': FlopsLoss(
    (model): SparseEncoder(
      (0): MLMTransformer({'max_seq_length': 256, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
      (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': None})
    )
  ), 'query_regularizer': FlopsLoss(
    (model): SparseEncoder(
      (0): MLMTransformer({'max_seq_length': 256, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
      (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': None})
    )
  )}
  

Evaluation Dataset

gooaq

• Dataset: gooaq at b089f72
• Size: 1,000 evaluation samples
• Columns: question and answer
• Approximate statistics based on the first 1000 samples:

  	question	answer
  type	string	string
  details	min: 8 tokens mean: 11.88 tokens max: 22 tokens	min: 14 tokens mean: 61.03 tokens max: 127 tokens

• Samples:

  question	answer
  how do i program my directv remote with my tv?	['Press MENU on your remote.', 'Select Settings & Help > Settings > Remote Control > Program Remote.', 'Choose the device (TV, audio, DVD) you wish to program. ... ', 'Follow the on-screen prompts to complete programming.']
  are rodrigues fruit bats nocturnal?	Before its numbers were threatened by habitat destruction, storms, and hunting, some of those groups could number 500 or more members. Sunrise, sunset. Rodrigues fruit bats are most active at dawn, at dusk, and at night.
  why does your heart rate increase during exercise bbc bitesize?	During exercise there is an increase in physical activity and muscle cells respire more than they do when the body is at rest. The heart rate increases during exercise. The rate and depth of breathing increases - this makes sure that more oxygen is absorbed into the blood, and more carbon dioxide is removed from it.

• Loss: SpladeLoss with these parameters:

  {'loss': SparseMultipleNegativesRankingLoss(
    (model): SparseEncoder(
      (0): MLMTransformer({'max_seq_length': 256, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
      (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': None})
    )
    (cross_entropy_loss): CrossEntropyLoss()
  ), 'lambda_corpus': 3e-05, 'lambda_query': 5e-05, 'corpus_regularizer': FlopsLoss(
    (model): SparseEncoder(
      (0): MLMTransformer({'max_seq_length': 256, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
      (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': None})
    )
  ), 'query_regularizer': FlopsLoss(
    (model): SparseEncoder(
      (0): MLMTransformer({'max_seq_length': 256, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
      (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': None})
    )
  )}
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• learning_rate: 2e-05
• num_train_epochs: 1
• bf16: True
• load_best_model_at_end: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 2e-05
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 1
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.0
• warmup_steps: 0
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 42
• data_seed: None
• jit_mode_eval: False
• use_ipex: False
• bf16: True
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_prefetch_factor: None
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: True
• ignore_data_skip: False
• fsdp: []
• fsdp_min_num_params: 0
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• fsdp_transformer_layer_cls_to_wrap: None
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: adamw_torch
• optim_args: None
• adafactor: False
• group_by_length: False
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: None
• hub_always_push: False
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• include_for_metrics: []
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• dispatch_batches: None
• split_batches: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: False
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• use_liger_kernel: False
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	Validation Loss	NanoClimateFEVER_dot_ndcg@10	NanoDBPedia_dot_ndcg@10	NanoFEVER_dot_ndcg@10	NanoFiQA2018_dot_ndcg@10	NanoHotpotQA_dot_ndcg@10	NanoMSMARCO_dot_ndcg@10	NanoNFCorpus_dot_ndcg@10	NanoNQ_dot_ndcg@10	NanoQuoraRetrieval_dot_ndcg@10	NanoSCIDOCS_dot_ndcg@10	NanoArguAna_dot_ndcg@10	NanoSciFact_dot_ndcg@10	NanoTouche2020_dot_ndcg@10	NanoBEIR_mean_dot_ndcg@10
0.0213	4000	0.3968	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.0425	8000	0.054	0.0224	0.2847	0.5628	0.8027	0.3260	0.6627	0.5252	0.3028	0.5467	0.7301	0.2563	0.3150	0.5072	0.4771	0.4846
0.0638	12000	0.0468	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.0850	16000	0.0394	0.0137	0.1908	0.5269	0.7778	0.3464	0.6510	0.5374	0.3086	0.5719	0.7901	0.2900	0.3661	0.5473	0.4839	0.4914
0.1063	20000	0.035	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1275	24000	0.0402	0.0142	0.1971	0.5098	0.6363	0.3715	0.6979	0.5442	0.3555	0.5223	0.7881	0.3008	0.3401	0.5963	0.4795	0.4877
0.1488	28000	0.0286	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1700	32000	0.0289	0.0209	0.2097	0.5169	0.7501	0.3622	0.6629	0.5151	0.3239	0.5322	0.8189	0.3121	0.3045	0.5318	0.4748	0.4858
0.1913	36000	0.0241	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2125	40000	0.0243	0.0166	0.2150	0.4990	0.6614	0.3184	0.6564	0.5499	0.2924	0.5506	0.8177	0.2755	0.3214	0.5292	0.4605	0.4729
0.2338	44000	0.021	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2550	48000	0.0205	0.0045	0.2210	0.5328	0.5836	0.3180	0.6990	0.5365	0.2860	0.5529	0.8704	0.2860	0.4025	0.6107	0.4314	0.4870
0.2763	52000	0.0181	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2975	56000	0.018	0.0129	0.2131	0.5543	0.7181	0.3645	0.6852	0.5199	0.3232	0.5970	0.8914	0.2980	0.4618	0.5037	0.4592	0.5069
0.3188	60000	0.0176	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3400	64000	0.018	0.0141	0.2607	0.4594	0.7357	0.3597	0.6538	0.5082	0.3070	0.4944	0.8569	0.3252	0.4125	0.5243	0.4489	0.4882
0.3613	68000	0.016	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3825	72000	0.0143	0.0082	0.2737	0.5459	0.7570	0.3845	0.6806	0.5035	0.3408	0.5338	0.8608	0.2888	0.3096	0.6163	0.4709	0.5051
0.4038	76000	0.0148	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4250	80000	0.0135	0.0211	0.2267	0.4964	0.7829	0.3579	0.6758	0.4954	0.3195	0.5164	0.8698	0.2745	0.3012	0.6260	0.4426	0.4912
0.4463	84000	0.0132	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4675	88000	0.012	0.0270	0.2442	0.5741	0.8005	0.3372	0.7019	0.5064	0.3109	0.6238	0.8988	0.2805	0.3875	0.5590	0.4396	0.5126
0.4888	92000	0.0126	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5100	96000	0.0127	0.0201	0.2948	0.5384	0.7822	0.3800	0.6947	0.5237	0.3674	0.5646	0.8843	0.2873	0.3825	0.5898	0.4812	0.5208
0.5313	100000	0.0113	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5525	104000	0.0112	0.0057	0.2318	0.5091	0.8362	0.3649	0.6829	0.4695	0.3442	0.5403	0.8920	0.2696	0.3787	0.6109	0.4384	0.5053
0.5738	108000	0.0094	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5951	112000	0.0095	0.0101	0.2325	0.5184	0.7349	0.3672	0.6673	0.4474	0.3196	0.5647	0.8866	0.2938	0.3345	0.5744	0.4609	0.4925
0.6163	116000	0.0096	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6376	120000	0.01	0.0084	0.2362	0.4989	0.8299	0.3595	0.6820	0.5200	0.3286	0.6138	0.8959	0.3088	0.4139	0.5808	0.4833	0.5194
0.6588	124000	0.0103	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6801	128000	0.0082	0.0115	0.2402	0.5127	0.7943	0.3828	0.6796	0.4925	0.3337	0.5848	0.8956	0.2880	0.3962	0.5981	0.4634	0.5124
0.7013	132000	0.0085	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7226	136000	0.0087	0.0125	0.2444	0.5258	0.7659	0.3397	0.6939	0.4942	0.3330	0.5573	0.8866	0.2789	0.3829	0.5305	0.4699	0.5002
0.7438	140000	0.0092	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7651	144000	0.0084	0.0071	0.2376	0.5247	0.8359	0.3551	0.6987	0.4440	0.3230	0.5973	0.8875	0.3052	0.4243	0.5601	0.4865	0.5138
0.7863	148000	0.0082	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8076	152000	0.0073	0.0036	0.2379	0.5045	0.8240	0.3389	0.7027	0.4895	0.3373	0.5893	0.8878	0.2870	0.3998	0.5728	0.4735	0.5112
0.8288	156000	0.0069	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8501	160000	0.0076	0.0024	0.2594	0.5254	0.8284	0.3557	0.6951	0.4858	0.3519	0.5968	0.8923	0.3212	0.439	0.6144	0.4877	0.5272
0.8713	164000	0.0062	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8926	168000	0.0061	0.0084	0.2580	0.5068	0.8307	0.3629	0.7095	0.5132	0.3373	0.5577	0.8803	0.3041	0.4438	0.5802	0.4668	0.5193
0.9138	172000	0.0067	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9351	176000	0.0072	0.0076	0.2627	0.4988	0.8192	0.3587	0.7072	0.4968	0.3488	0.5746	0.8794	0.3049	0.4671	0.5872	0.4739	0.5215
0.9563	180000	0.0049	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9776	184000	0.0056	0.0067	0.2672	0.4954	0.8207	0.3473	0.7148	0.4997	0.3479	0.5798	0.8778	0.3115	0.4557	0.5884	0.4753	0.5216
0.9988	188000	0.005	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
-1	-1	-	-	0.2594	0.5254	0.8284	0.3557	0.6951	0.4858	0.3519	0.5968	0.8923	0.3212	0.4390	0.6144	0.4877	0.5272

• The bold row denotes the saved checkpoint.

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 2.656 kWh
• Carbon Emitted: 1.032 kg of CO2
• Hours Used: 9.368 hours

Training Hardware

• On Cloud: No
• GPU Model: 1 x NVIDIA GeForce RTX 3090
• CPU Model: 13th Gen Intel(R) Core(TM) i7-13700K
• RAM Size: 31.78 GB

Framework Versions

• Python: 3.11.6
• Sentence Transformers: 4.2.0.dev0
• Transformers: 4.49.0
• PyTorch: 2.6.0+cu124
• Accelerate: 1.5.1
• Datasets: 2.21.0
• Tokenizers: 0.21.1

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

SpladeLoss

@misc{formal2022distillationhardnegativesampling,
      title={From Distillation to Hard Negative Sampling: Making Sparse Neural IR Models More Effective},
      author={Thibault Formal and Carlos Lassance and Benjamin Piwowarski and Stéphane Clinchant},
      year={2022},
      eprint={2205.04733},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2205.04733},
}