Awesome

Unified Multimodal Transformer (UMT) for Multimodal Named Entity Recognition (MNER)

Author

Jianfei Yu

July 1, 2020

The preprocessed CoNLL format files are provided in this repo. For each tweet, the first line is its image id, and the following lines are its textual contents.
Step 1：Download each tweet's associated images via this link (https://drive.google.com/file/d/1PpvvncnQkgDNeBMKVgG2zFYuRhbL873g/view)
Step 2: Change the image path in line 552 and line 554 of the "run_mtmner_crf.py" file
Step 3: Download the pre-trained ResNet-152 via this link (https://download.pytorch.org/models/resnet152-b121ed2d.pth)
Setp 4: Put the pre-trained ResNet-152 model under the folder named "resnet"

This is the training code of tuning parameters on the dev set, and testing on the test set. Note that you can change "CUDA_VISIBLE_DEVICES=2" based on your available GPUs.

sh run_mtmner_crf.sh

We show our running logs on twitter-2015 and twitter-2017 in the folder "log files". Note that the results are a little bit lower than the results reported in our paper, since the experiments were run on different servers.

In our codes, we mainly use "seqeval" to compute Micro-F1 as the evaluation metrics. Note that if you use the latest version of seqeval (as it may also report the weighted F1 score), you may need to change our Micro-F1 score parsing code as follows: float(report.split('\n')[-3].split(' ')[-2].split(' ')[-1]) to float(report.split('\n')[-4].split(' ')[-2].split(' ')[-1]).
In addition to "seqeval", we also borrow the evaluation code from this repo to compute Micro-F1. The Micro-F1 scores based on these two codes should be the same.

Using these two datasets means you have read and accepted the copyrights set by Twitter and dataset providers.
Most of the codes are based on the codes provided by huggingface: https://github.com/huggingface/transformers.