Hercules is a tool for checking code plagiarism in Github.
It can handle all sorts of changes in the code, and is also variable-name invariant. If there's an X% change, it will be able to tell you that there's a (100-X)% similarity.
git clone https://github.com/ongteckwu/hercules.git
cd hercules
//copy .env.example into .env
// then add your own GITHUB_TOKEN= into .env
Build the application:
./build.sh
Run the application:
./hercules --url=https://github.com/xxx/yyyy
OR
./hercules --dir=<path-to-code-directory>
Note: The application will take a couple of mins to run, due to the sheer volume of code to scan, and also to Github API limits.
1️⃣ Randomly picks N=15 code files from the assignment (ignores files that aren't code and folders).
2️⃣ Next, it uses token-level TFIDF to identify key terms in the code.
These terms are used to search GitHub for similar code files.
The top 10 best matched Github files are pulled per assignment code file.
3️⃣ Next, it applies two different methods, Double-side Argmin Levenshtein (DAL) and Char-level non-alphabet TFIDF (CLNAT), to see how similar the assignment code file is to the GitHub code files. 🔍
DAL: A form of Levenshtein that is used to find the most similar substring of string 1 in another string 2 (argmin). It returns:
- the substring correction count (the min) and
- the index of where the substring ends (the argmin).
- Using it on the reverse of string 1 and string 2 gives the index of where the substring starts (double-sided).
- The similarity score is
score = 1 - min/(sub_string_char_count)
CLNAT: This is TFIDF but on a character level. It ignores alphabets so that it is variable-name-change invariant.
4️⃣ It then counts the number of Github repositories that have similar code files.
Then, it picks the top M=8 similar repositories and compares them directly to the assignment using both DAL and CLNAT.
This results in three scores: summed weighted DAL, summed weighted CLNAT, and a weighted Combined Similarity.
- Weighted by character count e.g.
sum(character_count[i]/total_character_count * similarity_score[i])
combined_similarity_score = dal_score * clnat_score
5️⃣ Finally, it ranks these GitHub repositories based on the combined similarity score and shows the results in a table.
Note: N and M can be tuned.
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.
This project is licensed under the MIT License - see the LICENSE.md file for details.