Reports from the Heart of the Machine
Edit: Egg sauce
Academic misconduct in the field of AI has come again, an AAAI 2021 paper is suspected of plagiarizing the ACL 2020 paper, and some people have also posted the results of the double check.
Over the past few months, misconduct in the academic circle has been repeatedly exposed, first the ICCV paper of the master's degree of the University of Science and Technology of Hong Kong is suspected of plagiarism, and then there are "word for word" plagiarism of the top committee of the Beijing Polytechnic Master's Degree; not long ago, a blogger of station B posted a video saying that the key laboratory of Fudan University was suspected of plagiarizing the papers of American professors, and listed a series of evidence of plagiarism.
Recently, another netizen broke out that the AAAI 2021 and ACL 2020 papers are almost exactly the same: "The text part is almost wordlessly repeated."
The paper suspected of plagiarism is a paper from AAAI 2021 called Multi-modal Graph Fusion for Named Entity Recognition with Targeted Visual Guidance. The authors are from Southeast University and Soochow University.
Address of the paper: https://www.aaai.org/AAAI21Papers/AAAI-2753.ZhangD.pdf
The above paper is suspected of plagiarizing a paper from ACL 2020, "A Novel Graph-based Multi-modal Fusion Encoder for Neural Machine Translation", written by research institutions such as Xiamen University.
Address of the paper: https://arxiv.org/pdf/2007.08742.pdf
Let's take a look at what the two papers have in common.
Comparison of the contents of the two papers
First of all, there are similarities between the abstract and the introduction part of the paper, the following is a screenshot of the abstract of the two papers, through a simple comparison, we can find that there are a large number of similar sentences in the abstract, but a few keywords are replaced:
AAAI 2021 Suspected of plagiarizing papers
ACL 2020 Plagiarized Papers
In the main contribution part of the paper, the paper suspected of plagiarism has only been briefly modified:
There are also a number of similarities in the algorithm:
In order to more intuitively compare the two papers, someone uploaded a comparison of the double check results of the two papers: in the figure below, the ACL 2020 plagiarized paper on the left and the AAAI 2021 suspected plagiarized paper on the right. The results showed that there were 1.4K similar words in the two papers, 13.4% had the same part, 4.4% had small changes, and 5.3% had relevant meanings.
Comparison of check results. Source: Zhihu @sonta.
AAAI 2021 Author Response: Genus reference
The plagiarism incident has triggered a lot of discussion in the Zhihu community, and last night, one of the authors of the paper suspected of plagiarism, Zhihu user @Wei Suzhong, gave a positive response:
In response, the authors of the paper made several points:
The paper explicitly cites A Novel Graph-based Multi-modal Fusion Encoder for Neural Machine Translation
Related work also mentions connections and distinctions
Some of the code described in the github repository is referenced from acl2020
In addition, the authors say that the ACL 2020 paper is doing multimodal machine translation, and the AAAI 2021 paper is doing multimodal named entity recognition. There is no "plagiarism" relationship between the two papers.
"The method section refers to the paper from ACL2020. So there will be some similarities in the contribution and method sections. But in other parts, such as motivation, introduction, graph construction, experimentation, they are designed for our mner tasks."
At the same time, the author also said that the module will be re-edited and then resubmitted to AAAI Proceedings.
The author mentioned that the code for the AAAI 2021 paper has been open sourced and the experimental results can be reproduced. However, some people reported that the results of the experiment could not be reproduced according to the code provided in their GitHub, and the author did not reply to the issue.
What do you think about this?
Based on Python, the vehicle information identification system is quickly built using NVIDIA TAO Toolkit and Deepstream
NVIDIA TAO Toolkit is an AI toolkit that provides an off-the-shelf interface to ai/DL frameworks to build models faster without the need for coding.
DeepStream is a streaming media analysis toolkit for building AI applications. It takes streaming data as input and uses artificial intelligence and computer vision to understand the environment, converting pixels into data.
The DeepStream SDK can be used to build vision application solutions for traffic and pedestrian understanding in smart cities, health and safety monitoring in hospitals, self-inspection and analysis in retail, component defect detection in manufacturing plants, and more
December 14, 19:30-21:00, the summary of this sharing is as follows:
Introduces the latest features of tao toolkit;
Introduces the latest features of NVIDIA Deepstream;
Use TAO Toolkit's rich library of pre-trained models to quickly train models;
Deploy applications directly from TAO Toolkit's pre-trained models and Deepstream;
Complete the detection and identification of vehicle license plates, and detect the brand, color, and type of pedestrians and vehicles.