Shulou(Shulou.com)06/03 Report--

Produced by big data Digest

Compilation: Cao Peixin, Zhou Jiale

With the graduation season approaching, Zhai Tianlin was once again scolded by netizens on Weibo hot search.

After the plagiarism incident of Zhai's thesis, many colleges and universities have strengthened the paper evaluation standards, and some even have to check the thesis conclusion of non-graduates.

Many late-night paper-correcting students then came to Weibo @ Zhai Tianlin: are you asleep? How can you sleep? I'm still correcting my paper! Do you deserve to sleep?!

The picture is from Weibo

Zhai Tianlin himself may not have thought that he had made an outstanding contribution to China's higher education by mistake.

However, some low achiever said that plagiarism was impossible because the inspection was so strict in the future, so he had to find someone to write it for him.

Some media have exposed the industrial chain of "thesis writing" in academic circles. According to China Business report, the number of undergraduate liberal arts papers searched on Taobao is about 200 yuan per thousand words.

However, this "academic crooked road" is also going to be blocked by AI. According to a new study by researchers at the University of Copenhagen, a "anti-shooter ghostwriting" AI system has just been developed.

The system is intended to detect paper cheating through intelligent writing analysis technology. According to your writing habits, you can determine whether the paper is written by yourself or by someone else.

Based on an analysis of 130000 written assignments, scientists can detect with nearly 90 per cent accuracy whether students wrote the homework themselves or by ghostwriters.

Basically, we can achieve the accurate result of "ghostwriting" homework.

Writing in senior high school is popular, the divine operation of the University of Copenhagen.

The study of cheating on students' papers has been carried out at the Department of computer Science (DIKU) at the University of Copenhagen for several years, initially for high school students in Denmark.

In Denmark, the homework checking platform currently mainly used in high schools is called Lectio, which can be used to check whether any paragraphs in students' assignments are directly copied from previously submitted assignments.

However, with the popularity of various online service platforms, it is becoming easier for Danish high school students to find someone to do their homework.

In the face of this situation, the school has been lack of effective testing means.

The Learning track Program or "SRP" (Danish for "Studie Retnings Projekt") is a required interdisciplinary course for Danish high school graduates and a very important written assignment. The phenomenon of cheating in this project is particularly striking.

Because SRP is so important for graduation, many students post their writing assignments on the Danish auction website Den Bla Avis to find someone to write it.

Like Chinese teachers and most weight-checking systems, Lectio can only check duplicates and cannot determine whether an assignment is written by someone else.

Some departments of the University of Copenhagen have been working with many high schools on SRP programs, which suffer from ghostwriting cheating and have been exploring solutions.

The DABAI project team of the university's computer science department decided to teach these lazy high school students to be "human beings".

DABAI (Danish big data Analysis-driven Innovation Center) is a Danish national research center established in 2016. In addition to studying efficient algorithms for machine learning, the research group has paid special attention to student education. Previously, they have studied educational projects such as "optimizing students' personalized learning" and "improving teachers' insight".

An anti-gunner artifact called "the shooter."

This cheating prevention program, called Ghostwriter, is essentially a text analysis program based on machine learning and neural network technology.

Dr. Stephan Lorenzen, a member of the project team, said the program can compare the student's recently submitted and previously submitted articles to identify differences in writing styles.

"the program will pay attention to many features such as the length of the word, the sentence structure, and the way the word is used. For example, it will detect whether 'for example' is written as' ex', or 'e.g.journal'."

Its data set comes from MaCom, which provides a Lectio platform for Danish high schools, which covers more than 90 per cent of Danish high schools and provides 130000 written assignments for students from different high schools to researchers on the GhostWriter project.

The research team believes that the product is very practical, and that many schools have a growing technical demand to find out who wrote the paper.

But Dr. Stephan Lorenzen also believes that "before that, there is a need to seriously discuss the ethical issues of using this technology. We should not take the conclusions of this procedure as the only criterion for cheating, but should regard it as supporting evidence."

How does Ghostwriter work?

The Ghostwriter program uses Siamese neural network to distinguish the writing styles of different texts: through the training of a large amount of data, we learn the external performance (representation) of different writing styles, and then compare them.

This project solves the author authentication problem in two steps. First of all, the problem of calculating the similarity of writing style between two texts is solved, mainly by using the Siamese network learning similarity function SJV T × T → [0Power1]. The second is to solve the verification problem of author A by comparing the similarity between the unknown author text X and the text T known to be author A.

On the network side, they considered using different input channels to consider several different architectures (for example, char,word,POS-tags) and finally determined a network architecture that performed best:

Best performing network

The encoding part includes a character embedding (Embd), followed by two unused convolution layers, each followed by a global maximum pooling layer (GMP).

In the comparison part, they first calculate the absolute difference between the codes in the merge layer, then apply four dense layers with 500neurons in each layer, and finally use the softmax layer with two outputs for normalization.

They divided the data set into three parts: T-train for training, T-val for early stop of training and selecting Cs,T-test for estimating test models only.

After training, the accuracy of the model reached 87.5%.

The final function is that when a student submits an assignment, the network compares it with the previous assignment. For each job, the neural network calculates a percentage to represent the similarity between the old and the new jobs. Then, a weighted average is calculated by comprehensively considering many factors, such as the similarity of new and old homework, the time of handing in homework and so on. This final value can be used to show the similarity between the new assignment and the students' writing style.

The study has been published in a paper called "identifying the Gunners in High School."

Links to papers:

Https://www.science.ku.dk/presse/nyhedsarkiv/2019/fristet-til-at-snyde-med-eksamensopgaven-kunstig-intelligens-opdager-dig-med-90-procent-sikkerhed/Detecting_Ghostwriters_in_High_Schools.pdf

In addition to papers, you can also cooperate with the police to screen forged texts.

In addition to ghosting homework, Ghostwriter's technology can also be applied to other parts of society.

For example, the program can assist police document examiners to analyze the authenticity of various documents, such as whether a business contract is forged, or whether the deceased left a suicide note in a bizarre suicide case. whether the suicide note was written by the deceased himself, and so on.

"it will be interesting to cooperate with the police. the existing method for the police is to hire document examiners to qualitatively compare the similarities and differences between texts. Our method is applicable to big data and automatically find the hidden patterns. I think the combination of the two will be beneficial to the police." Lorenzen said he stressed the need to discuss the ethical issues he faces here as well.

This technology, which uses artificial intelligence to detect cheating in homework, has a wide application prospect.

Currently, it is also used to analyze Twitter text to determine whether the text is written by real users or by water armies or robots. In other words, Taobao stores hire water army praise, it is likely to be identified.

Related reports:

Https://www.sciencedaily.com/releases/2019/05/190529145048.htm

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.