metadata

license: mit
datasets:
  - philipphager/baidu-ultr-pretrain
  - philipphager/baidu-ultr_uva-mlm-ctr
metrics:
  - dcg@1
  - dcg@3
  - dcg@5
  - dcg@10
  - ndcg@10
  - mrr@10

Naive Listwise MonoBERT trained on Baidu-ULTR

A flax-based MonoBERT cross encoder trained on the Baidu-ULTR dataset with a listwise softmax cross-entropy loss on clicks. The loss is called "naive" as we use user clicks as a signal of relevance without any additional position bias correction. For more info, read our paper and find the code for this model here.

Test Results on Baidu-ULTR Expert Annotations

Model	log-likelihood	DCG@1	DCG@3	DCG@5	DCG@10	nDCG@10	MRR@10
Pointwise Naive	0.2272	1.6836	3.5616	4.8822	7.4244	0.3640	0.6096
Pointwise Two Tower	0.2178	1.4826	3.2636	4.5491	7.0979	0.3476	0.5856
Pointwise IPS	0.2436	0.8842	2.0510	2.9535	4.8816	0.2363	0.4472
Listwise Naive	-	1.9738	4.1609	5.6861	8.5432	0.4091	0.6436
Listwise IPS	-	1.7466	3.6378	4.9797	7.5790	0.3665	0.6112

Usage

Here is an example of downloading the model and calling it for inference on a mock batch of input data. For more details on how to use the model on the Baidu-ULTR dataset, take a look at our training and evaluation scripts in our code repository.

import jax.numpy as jnp

from src.model import ListwiseCrossEncoder

model = ListwiseCrossEncoder.from_pretrained(
    "philipphager/baidu-ultr_uva-bert_naive-listwise",
)

# Mock batch following Baidu-ULTR with 4 documents, each with 8 tokens
batch = {
    # Query_id for each document
    "query_id": jnp.array([1, 1, 1, 1]),
    # Document position in SERP
    "positions": jnp.array([1, 2, 3, 4]),
    # Token ids for: [CLS] Query [SEP] Document
    "tokens": jnp.array([
        [2, 21448, 21874, 21436, 1, 20206, 4012, 2860],
        [2, 21448, 21874, 21436, 1, 16794, 4522, 2082],
        [2, 21448, 21874, 21436, 1, 20206, 10082, 9773],
        [2, 21448, 21874, 21436, 1, 2618, 8520, 2860],
  ]),
    # Specify if a token id belongs to the query (0) or document (1)
    "token_types": jnp.array([
        [0, 0, 0, 0, 1, 1, 1, 1],
        [0, 0, 0, 0, 1, 1, 1, 1],
        [0, 0, 0, 0, 1, 1, 1, 1],
        [0, 0, 0, 0, 1, 1, 1, 1],
    ]),
    # Marks if a token should be attended to (True) or ignored, e.g., padding tokens (False):
    "attention_mask": jnp.array([
        [True, True, True, True, True, True, True, True],
        [True, True, True, True, True, True, True, True],
        [True, True, True, True, True, True, True, True],
        [True, True, True, True, True, True, True, True],
    ]),
}

outputs = model(batch, train=False)
print(outputs)

Reference

@inproceedings{Hager2024BaiduULTR,
  author = {Philipp Hager and Romain Deffayet and Jean-Michel Renders and Onno Zoeter and Maarten de Rijke},
  title = {Unbiased Learning to Rank Meets Reality: Lessons from Baidu’s Large-Scale Search Dataset},
  booktitle = {Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR`24)},
  organization = {ACM},
  year = {2024},
}