CCF BDCI 2020 房产行业聊天问答匹配赛道 A榜47/2985

Overview

CCF BDCI 2020 房产行业聊天问答匹配 A榜47/2985

赛题描述详见:https://www.datafountain.cn/competitions/474

文件说明

data: 存放训练数据和测试数据以及预处理代码

model_bert.py: 网络模型结构定义

adv_train.py: 对抗训练代码

run_bert_pse_adv.py: 运行bert-wwm + 对抗训练 + 伪标签模型

run_roberta_cls_pse_reinit_adv.py: 运行roberta-large last2embedding_cls + reinit + 对抗训练 + 伪标签模型

个人方案

我的baseline是将query和answer拼接后传入预训练好的bert进行特征提取,之后将提取的特征传入一个全连接层,最后接一个softmax进行分类。

其中尝试的预训练模型有bert(谷歌),bert_wwm(哈工大版本),roberta_large(哈工大版本),xlneternie等,其中效果较好的有bert-wwm和roberta-large。之后在baseline的基础上进行了各种尝试,主要尝试有以下:

模型 线上F1
bert-wwm 0.78
bert-wwm + 对抗训练 0.783
bert-wwm + 对抗训练 + 伪标签 0.7879
roberta-large 0.774
roberta-large + reinit + 对抗训练 0.786
roberta-large + reinit+对抗训练 + 伪标签 0.7871
roberta-large last2embedding_cls + reinit + 对抗训练 + 伪标签 0.7879

对抗训练

其基本的原理呢,就是通过添加扰动构造一些对抗样本,放给模型去训练,以攻为守,提高模型在遇到对抗样本时的鲁棒性,同时一定程度也能提高模型的表现和泛化能力。

参考链接:https://zhuanlan.zhihu.com/p/91269728

伪标签

将测试数据和预测结果进行拼接,之后当成训练数据传入到模型中重新进行训练。为了减少对训练数据的原始分布的影响并增加伪标签的置信度,我只在五个采用不同预训练模型的baseline预测一致的数据中采样了6000条测试数据加入到训练集进行训练。

重新初始化

参考链接:如何让Bert在finetune小数据集时更“稳”一点 https://zhuanlan.zhihu.com/p/148720604

大致思想是靠近底部的层(靠近input)学到的是比较通用的语义方面的信息,比如词性、词法等语言学知识,而靠近顶部的层会倾向于学习到接近下游任务的知识,对于预训练来说就是类似masked word prediction、next sentence prediction任务的相关知识。当使用bert预训练模型finetune其他下游任务(比如序列标注)时,如果下游任务与预训练任务差异较大,那么bert顶层的权重所拥有的知识反而会拖累整体的finetune进程,使得模型在finetune初期产生训练不稳定的问题。

因此,我们可以在finetune时,只保留接近底部的bert权重,对于靠近顶部的层的权重,可以重新随机初始化,从头开始学习。

在本次比赛中,我只对最后roberta-large的最后五层进行重新初始化。在实验中,我发现对于bert,重新初始化会降低效果,而roberta-large则有提升。

bert 不同embedding和cls组合

思路主要是参考 CCF BDCI 2019 互联网新闻情感分析 复赛top1解决方案

参考链接:https://github.com/cxy229/BDCI2019-SENTIMENT-CLASSIFICATION

即对bert不同embedding进行组合后传入全连接层进行分类。该方案尝试时间较晚,只实验last2embedding_cls这种组合,结果也确实有提升。

模型融合

对于单模,我采用五折交叉验证,对每一个单模的五个模型结果,我尝试了相加融合和投票的方式,结果是融合相加的线上f1较高

对于不同模型,我也只是采用的相加融合的方式(由于时间问题没有尝试投票和stacking的方式)。最后a榜效果最好的是bert-wwm + 对抗训练 + 伪标签、roberta-large + reinit+对抗训练 + 伪标签、roberta-large last2embedding_cls + reinit + 对抗训练 + 伪标签 三个模型的融合,线上F1有 0.7908 , 排名47;B榜我尝试只对两个效果最好的模型进行融合,即 bert-wwm + 对抗训练 + 伪标签last2embedding_cls + reinit + 对抗训练 + 伪标签,最终F1为0.80,排名72。

总结

本次参加比赛完全是数据挖掘课程要求,也是我第一次参加大数据比赛。因为我的研究方向是图像,所以基本可以说是从零开始,写这个github只是想记录一下这一个月自己从零开始的参赛经历,也希望对同样参加类似比赛的新人有帮助。最后,希望看到了顺手给star,万分感谢。

Owner
shuo
shuo
Deeply Supervised, Layer-wise Prediction-aware (DSLP) Transformer for Non-autoregressive Neural Machine Translation

Non-Autoregressive Translation with Layer-Wise Prediction and Deep Supervision Training Efficiency We show the training efficiency of our DSLP model b

Chenyang Huang 37 Jan 04, 2023
Repository for Project Insight: NLP as a Service

Project Insight NLP as a Service Contents Introduction Features Installation Setup and Documentation Project Details Demonstration Directory Details H

Abhishek Kumar Mishra 286 Dec 06, 2022
This is a simple item2vec implementation using gensim for recbole

recbole-item2vec-model This is a simple item2vec implementation using gensim for recbole( https://recbole.io ) Usage When you want to run experiment f

Yusuke Fukasawa 2 Oct 06, 2022
To classify the News into Real/Fake using Features from the Text Content of the article

Hoax-Detector Authenticity of news has now become a major problem. The Idea is to classify the News into Real/Fake using Features from the Text Conten

Aravindhan 1 Feb 09, 2022
Application to help find best train itinerary, uses speech to text, has a spam filter to segregate invalid inputs, NLP and Pathfinding algos.

T-IAI-901-MSC2022 - GROUP 18 Gestion de projet Notre travail a été organisé et réparti dans un Trello. https://trello.com/b/X3s2fpPJ/ia-projet Install

1 Feb 05, 2022
Partially offline multi-language translator built upon Huggingface transformers.

Translate Command-line interface to translation pipelines, powered by Huggingface transformers. This tool can download translation models, and then us

Richard Jarry 8 Oct 25, 2022
The model is designed to train a single and large neural network in order to predict correct translation by reading the given sentence.

Neural Machine Translation communication system The model is basically direct to convert one source language to another targeted language using encode

Nishant Banjade 7 Sep 22, 2022
ttslearn: Library for Pythonで学ぶ音声合成 (Text-to-speech with Python)

ttslearn: Library for Pythonで学ぶ音声合成 (Text-to-speech with Python) 日本語は以下に続きます (Japanese follows) English: This book is written in Japanese and primaril

Ryuichi Yamamoto 189 Dec 29, 2022
Nested Named Entity Recognition for Chinese Biomedical Text

CBio-NAMER CBioNAMER (Nested nAMed Entity Recognition for Chinese Biomedical Text) is our method used in CBLUE (Chinese Biomedical Language Understand

8 Dec 25, 2022
GCRC: A Gaokao Chinese Reading Comprehension dataset for interpretable Evaluation

GCRC GCRC: A New Challenging MRC Dataset from Gaokao Chinese for Explainable Eva

Yunxiao Zhao 5 Nov 04, 2022
Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning

GenSen Learning General Purpose Distributed Sentence Representations via Large Scale Multi-task Learning Sandeep Subramanian, Adam Trischler, Yoshua B

Maluuba Inc. 309 Oct 19, 2022
GPT-2 Model for Leetcode Questions in python

Leetcode using AI 🤖 GPT-2 Model for Leetcode Questions in python New demo here: https://huggingface.co/spaces/gagan3012/project-code-py Note: the Ans

Gagan Bhatia 100 Dec 12, 2022
Guide to using pre-trained large language models of source code

Large Models of Source Code I occasionally train and publicly release large neural language models on programs, including PolyCoder. Here, I describe

Vincent Hellendoorn 947 Dec 28, 2022
Repo for Enhanced Seq2Seq Autoencoder via Contrastive Learning for Abstractive Text Summarization

ESACL: Enhanced Seq2Seq Autoencoder via Contrastive Learning for AbstractiveText Summarization This repo is for our paper "Enhanced Seq2Seq Autoencode

Rachel Zheng 14 Nov 01, 2022
Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing

Introduction Funnel-Transformer is a new self-attention model that gradually compresses the sequence of hidden states to a shorter one and hence reduc

GUOKUN LAI 197 Dec 11, 2022
Simple translation demo showcasing our headliner package.

Headliner Demo This is a demo showcasing our Headliner package. In particular, we trained a simple seq2seq model on an English-German dataset. We didn

Axel Springer News Media & Tech GmbH & Co. KG - Ideas Engineering 16 Nov 24, 2022
Code Implementation of "Learning Span-Level Interactions for Aspect Sentiment Triplet Extraction".

Span-ASTE: Learning Span-Level Interactions for Aspect Sentiment Triplet Extraction ***** New March 31th, 2022: Scikit-Style API for Easy Usage *****

Chia Yew Ken 111 Dec 23, 2022
SDL: Synthetic Document Layout dataset

SDL is the project that synthesizes document images. It facilitates multiple-level labeling on document images and can generate in multiple languages.

Sơn Nguyễn 0 Oct 07, 2021
Prompt-learning is the latest paradigm to adapt pre-trained language models (PLMs) to downstream NLP tasks

Prompt-learning is the latest paradigm to adapt pre-trained language models (PLMs) to downstream NLP tasks, which modifies the input text with a textual template and directly uses PLMs to conduct pre

THUNLP 2.3k Jan 08, 2023
Based on 125GB of data leaked from Twitch, you can see their monthly revenues from 2019-2021

Twitch Revenues Bu script'i kullanarak istediğiniz yayıncıların, Twitch'den sızdırılan 125 GB'lik veriye dayanarak, 2019-2021 arası aylık gelirlerini

4 Nov 11, 2021