File size: 3,222 Bytes
84c2e71 925ce6c 51fb168 1b6c7bf 79b73b4 51fb168 79b73b4 797f9d4 f770d24 51fb168 f770d24 59d07ef 797f9d4 59d07ef 797f9d4 59d07ef 797f9d4 59d07ef 797f9d4 59d07ef 797f9d4 51fb168 79b73b4 51fb168 1b6c7bf |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 |
---
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](https://arxiv.org/abs/2207.03051) 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](https://arxiv.org/abs/2404.02543) and [find the code for this model here](https://github.com/philipphager/baidu-bert-model).
## Test Results on Baidu-ULTR Expert Annotations
| Model | log-likelihood | DCG@1 | DCG@3 | DCG@5 | DCG@10 | nDCG@10 | MRR@10 |
|---------------------|----------------|--------|--------|--------|--------|---------|--------|
| Naive Pointwise | 0.2272 | 1.6836 | 3.5616 | 4.8822 | 7.4244 | 0.3640 | 0.6096 |
| **Naive Listwise** | - | 1.9738 | 4.1609 | 5.6861 | 8.5432 | 0.4091 | 0.6436 |
## Usage
Here is an example with a mock input batch for how to download and call the model:
```Python
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)
print(outputs)
```
For more details on how to use the model with real data from Baidu-ULTR, take a look at the [evaluation script of our model repository](https://github.com/philipphager/baidu-bert-model/blob/main/eval.py).
## 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},
}
```
|