【Solution ZhichunRoad】 5th@task1, 7th@task2 and 8th@task3

First of all, many thanks to the organizers for their hard work(the AICrowd platform& the Shopping Queries Dataset[1]).
And I learned a lot from the selfless sharing of the contestants in the discussion board.
Below is my solution:

|  SubTask  |         Methods           |       Metric      |  Ranking  |
|   task1   |  ensemble 6 large models  |    ndcg=0.9025    |    5th    |
|   task2   |    only 1 large model     |  micro f1=0.8194  |    7th    |
|   task3   |    only 1 large model     |  micro f1=0.8686  |    8th    |

It seems to me that this competition mainly contains two challenges:

Q1.How to improve the search quality of those unseen queries?

A1: we need more general encoded representations.

Q2.There is very rich text information on the product side, how to fully characterize it ?

A2: As the bert-like model’s “max_lenth paramter” increases, the training time increases more rapidly.
We need an extra semantic unit to cover all the text infomation.

A1 Solution:
Inspired by some multi-task pre-training work[2]
In pre-training stage, we adopt mlm task, classification task and contrastive learning task to achieve considerably performance.


A2 Solution
Inspired by some related work[3,4]
In fine-tuning stage, we use confident learning, exponential moving average method (EMA), adversarial training (FGM) and regularized dropout strategy (R-Drop) to improve the model’s generalization and robustness.
Moreover, we use a multi-granular semantic unit to discover the queries and products textual metadata for enhancing the representation of the model.

image image

In this work, we use data augmentation, multi-task pre-training and several fine-tuning methods to imporve our model’s generalization and robustness.

We release the source code at GitHub - cuixuage/KDDCup2022-ESCI

Thanks All~ :slight_smile:
Looking forward to more sharing and papers

[1] Shopping Queries Dataset: A Large-Scale ESCI Benchmark for Improving Product Search
[2] Multi-Task Deep Neural Networks for Natural Language Understanding
[3] Embedding-based Product Retrieval in Taobao Search
[4] Que2Search: Fast and Accurate Query and Document Understanding for Search at Facebook


Nice work! But in our experiment. pre-train with MLM task is useless. Bro, have you done any vigorous ablation experiments on this?

I have same problem about MLM task

MLM post-train didn’t work in our experiment as well.:thinking:

Thanks for the information, they are very helpful!

I used the mlm task as the basic pre-training task and did not remove it separately.
So I not sure whether the MLM Task works well.

But, I would like to provide some additional information.

  1. As presented in some Domain pre-training Papers[1],using additional unlabeled datasets for pretraining can get good performance in most domains CV, NLP, etc. (I personally think the pre-training stage is essential.)

  2. In the product_catalogue dataset, there are ~20w product meta info that do not appear in the public dataset (the training set & public test set).
    To make the model “familiar with” this part of the data. Except for the MLM task, I can’t think of a better method…

[1] Don’t Stop Pretraining: Adapt Language Models to Domains and Tasks

@yrquni @hsuchengmath @zhichao_feng
Looking forward to more discussions & ideas.

Thank your reply, dont stop pre-train, I think it is good idea.

If we have private dataset, then we can prove MLM workable.

Hi, mohanty
Please release private dataset.