TestPod

Automated Evaluation Of Essay-Based Questions Based On Natural Language Processing (AI)

Video Demonstration

One of the most crucial aspects of a teaching faculty is to correct the answer sheets
With a curriculum driven by tests and scoring, examinations have stood as the backbone of the education industry.
Teaching faculty spend an enormous amount of time for verification and assigning the scores to each student. The following action is a limitation as the faculty can focus on other aspects of teaching and enriching the students.
The application focuses on a novel methodology to automate the correction for subjective-based questions and answers using NLP Techniques.

The major stakeholders in the application are students and faculty. The faculty is capable of creating a test, monitoring the automated evaluation of each student, and seeing the statistics of the test. The student can logon to the application, attend the test by submitting the test ID, and submit the test for automated evaluation.
The model is built using Apache OpenNLP. An answer is decomposed into a set of sentences, and sentences are decomposed into tokens. The model decomposes the paragraph into sentences. The set of sentences in an answer goes through text transformation, which includes the removal of trailing spaces and case conversion. Following the transformation of the set of sentences, the text is fed as input to a part-of-speech tagger, which tags the tokens with the appropriate tag.
The mentioned process is performed on the faculty answer script to generate a skeleton that includes the expected answers. When a student submits an answer, the answer is transformed into a skeleton of tags and tokens, compared with the faculty answer script, and evaluated. The final score is calculated using a weighted average.

Purpose	Name
Web Framework	Spring MVC
Front-End	JSP & JSTL
Database	MYSQL
NLP	Apache Open NLP
Sentence Dedection	SentenceDetectorME (OpenNLP)
POS Tagger	en-pos-maxent (Maximum entropy Algorithm)
Remarks Generation	Rule-based fuzzy Logic

Purpose	Description
Home page	Contains information regarding the application
Login	Common login for faculty and student
Faculty Dashboard	Contains details about the faculty and the tests that were administered
Exam Dashboard	Test statistics and a list of students who took the test are projected
Auto Generated Report	Summary of the student's performance, step-by-step evaluation of the student report
Examination Centre	Evaluates the presence of the student's ID in the test and generates the question paper
Digital Question Paper	Platform for the student to answer the questions

The answer is broken down into sentences, using the delimiter (.)
The broken array of sentences is parsed using POS sentence tagging
The sentences are mapped to the token as <POS Tag, Token> mapping
For each sentence, the sentence is iterate through the resources provided to the model for analysis and evaluation
An intermediate output in the form of an array of boolean values are generated for further processing by rule-based NLP Techniques
The intermediate output is fed to a fuzzifier to generate fuzzy sets for discrete evaluation and using hierarchical-based valuation
Finally, the score for the submitted answer is generated using the weighted average technique from the discrete score from the fuzzy sets

The application is built using Apache OpenNLP, an NLP tool built for Java applications
The major stakeholder of the application are the faculty and the student. The faculty feeds the question and the respective answer key to the application
The uploaded questions and answers are processed respectively
Every test has a test ID, a unique primary key to map the question with the answer.
The questions for the test are saved in RDBMS. The answers are processed using the core NLP package in the application
The answer key is processed, tokenized, and POS tagged. The structured answer is stored in a text file along with the metadata of the file
When a student attempts the exam, the answers are processed, tokenized and POS tagged. The structured answer script saved as a file is decoded to proper data structures with the help of metadata
A container (Java bean) is created to save the information of the structure. Two containers, one for faculty and another for a student are created
The student container is passed to the faculty container to find the similarity index matrix
The matrix is de-fuzzified using fuzzy logic and appropriate marks are allocated for each question using weighted average method