NLP Grading and plagiarism - APIS

1. Setup

pip install -r requirments.txt

python src/models/model_download.py

python -m spacy download en_core_web_sm

2. Usage

2.1 API side

!python src/main.py

2.1.1 Plagiarsim

import requests as req


inp = {"essays_dict": 
   {"p7": "carpooled overtone censurers Gorey tyrannises",
   "uv8": "amphimixis Oozlefinch mesmerize attaman untrue",
   "7mgj": "remodelled unitholders toddled uncousinly pug-dog",
   "evjd2": "shortness tupara potato-moth proximates digger",
   "yDim":  "amphimixis mesmerize attaman untrue"
   },
  "cased": false}

out = req.post(url, json=inp)

print(out.json())

{
    "plagiarism_results": [
        {
            "uv8": {
                "yDim": 0.7832804918289185
            }
        },
        {
            "yDim": {
                "uv8": 0.7832803726196289
            }
        }
    ]
}

2.1.2 Grading

import requests as req

inp = {"essays_dict": {
        "78vv74": "\"\"@increase@\"\"1.0",
        "9": "name all is good hi gallon increase",
        "10": "all is good dafdr",
        "11": "all is decrease dafdr"
    },
    "grading_model": "manual_keywords_grading",
    "cased": false
    }

out = req.post(url, json=inp)

print(out.json())

{
    "grades": {
        "9": 1.0,
        "10": 0.0,
        "11": 0.0
    }
}

2.2 Individual classes

# has the encoder singleton
import transformer_model
# for grading
import keywords_grading

bert = transformer_model()
KGM = keywords_grading(bert)

answers = ["ahmed is eating a pizza",
        "i go to school by bus",
        "i am eating a pizza",
        "ahmed is eating a pizza"
        ]
ids = [i in range(len(answers))]

# call the pipeline
KGM.predict(answers= answers ,
        ids = ids,
        )

{
    "grades": {
        "9": 0.177,
        "10": 0.475,
        "11": 0.93
    }
}

3. Data

3.1 Grading

ASAP dataset

3.1 Plagiarism

quora question pairs

arXiV

4. Appindix

manual keywords grading model grammer

1) hard grading

enclosure is ""

# NOTE: all weights are optional
weight:float or None = 1.0

example = f"\"\"keyword-phrase\"\""{weight}

keywords are just matched with students answer if they contain them

2) soft grading

enclosure is ""@

weight:float or None = 1.0
example = f"\"\"@keyword-phrase@\"\"{weight}"

students answers are splited in keyword-phrases' n_grams. Then encoded and get maximum similarity between them and the keyword-phrases if the similarity value > SOFT_THRESHOLD put to one else zero for each keyword-phrase

3) medium grading

enclosure is ""@@

weight:float or None = 1.0
example = f"\"\"@@keyword-phrase@@\"\"{weight}"

students answers are splited in keyword-phrases' n_grams. Then encoded and get maximum similarity between them and the keyword-phrases if the similarity value > MEDUIM_THRESHOLD put to one else zero for each keyword-phrase

Note: that MEDUIM_THRESHOLD > SOFT_THRESHOLD

Note: that you can combine the three grading types interchangeably

example = "\"\"@@keyword_one@@\"\" \"\"keyword_two\"\" rest of the sentence"