Student testing and monitoring system (STMS) using Nlp

Published Online September 2019 in MECS DOI: 10.5815/ijmecs.2019.09.03

As technology has always been the most rapidly growing domain out of all others. From clocks to smart watches, calendars, or calculators; everything has renovated and taken a new compact form. The same is happening in the domain of knowledge and learning. There is an observed rising concern with the creation of such a platform that allows a common user, from any field to perform quizzes based on Multiple-Choice Questions(MCQs) questions from whatever domain they choose [1]. Certainly, it would also help those who lack in practical experience of freelancing but have theoretical knowledge.

NLP is being used for the automatic generation of questions. As educationists need to conduct the examination to judge student’s calibre and creating questions manually is very time-consuming so developing a system which automatically generates questionnaire will be of quite a huge help.

Furthermore, the automated process of generating questions on the premises of the course given, the system of generating questions are isolated into two units i.e. selection of required content and question production. Questions are formed by finding relevant parts of the text which constitutes the content selection. Researcher has seven discourse connectives in concentration i.e. like because, since, although, as a result, for example, and for instance, on that basis of which Question type will be decided [3]. As, if a sentence consists of “since” then question type would be “When”. The System has been evaluated for semantic and syntactic soundness of question by two evaluators. Such a system can also be called as a multi-agent system because of its working.

“Semantic Based Automatic Question Generation system uses two techniques named as Semantic Role Labeling and NER (Named Entity Recognizer) technique that converts the inputted sentence to a semantic pattern while developed and artificial immune system would be able to classify the patterns according to the question type” [4]. Feature extraction, learning, storage memory and associative retrieval are the features utilized by the immune system for the purpose of solving classification and recognition tasks. NER and SRL techniques will be used to parse inputted sentence which will specify whether there are person name, location, date in the sentence and through this identification a pattern for the question will be proposed, for example, if there is a person’s name mentioned then pattern of the question would be “Who” [5].

This technique of question generation gives a promising ratio of 87% of truly generated question patterns [4,6]. This ratio is far beyond other way of generating questions automatically.

It is researcher’s job to see whether the questions this system is producing can be used as pre-questions or not. Text-based and supporting image are the types of prequestions to be investigated.

The system suggested here further has three subsystems: Generator of Questions, Descriptor of Knowledge and Executer of E-Learning. The subsystem of Knowledge Descriptor allows the teacher to explain the learning contents. The subsystem of Questions Generator would receive the learning materials and produce required questions in multiple choice questions. The subsystem of E-learning Executer will utilize the generated questions and will let the students to use these questions in education process. While system generated correct and wrong options question (CWOQ) algorithm would be executed relying on sub-algorithm generating One solution, two Solutions, All of The Above, and none Of The Above algorithms.” A question bank have been developed for course General Biology course (Bio110) at Faculty of Science, King Abdul-Aziz University (KAU)” [6,7]. The bank contains 12 sections with 239 sub-points and a sum of 46345 inquiries. The student through this system gets the ability to solve the exercises of each section and the shows the given and correct answer after self-produced multiple-choice questions from domain ontologies [7].

Domain specific ontologies are the basics of the approach shown in this paper and lexicons such as WordNet or other linguistic resources aren’t utilized [8]. Semantic Web standard technology OWL (Ontology Web Language) is used by the system to generate multiplechoice questions. The System creates multiple choice question items using the Class-based, property-based, terminology-based strategies. Property-based strategies may produce a large number of multiple-choice questions but are very difficult to manipulate syntactically. The created tests were assessed in three dimensions: Pedagogical quality, phonetic/linguistic accuracy and the quantity of questions delivered.

Another research is performed to generate questions from documents written in the Punjabi language. An NER (Named Entity Recognition) Tool was needed to create which can recognize the nouns from a given phrase or sentence and produce proper questions from it. [9] Moreover, mind the Gap: Learning to Choose Gaps for Question Generation needs to address the problem of generating good questions into two parts: first, select sentences to be queried about, and second, to know which part of the sentence will be turned into a question. To achieve the goal of selecting better gap-fill questions, the author has broken down the task into stages i.e. Sentence selection, Question construction and Distractors generation [9,10].

Features brought into consideration for generating questions are Token count, lexical, syntactic, and semantic and NER feature is used to generate the fill in the blanks question [11]. The system would consist of a straightforward pipeline. First the source text is broken down into sentences to be processed by SENNA software, SENNA does the tokenizing, POS tagging, syntactic constituency parsing and semantic role labelling. A matcher function returns a list of patterns that match the predicate-argument structure of the sentences from the source. This function is called for each predicate and its linked semantic arguments [12]. Paper focuses on evaluating the produced questions basically in terms of their linguistic quality.

• II.  Related Works

CQG (Cloze Question generation) system is the system that produced a list of cloze questions in English article. CQG system is broken down into three main units, i.e. the selection of sentence, key and distractor. Initial stages involve the selection of relevant and informative portions of the text and keywords from the selected text are recognized out in the second stage [13]. Key selection will not be noun or adjective instead it will be found on the basis of NER. The third stage involves the selection of Distractors. First two stages are not bound in a fixed domain while the third one is bound. “Because quality of distractor depends on domain so distractor will be selected on the basis of the key selected and through web, list of distractors will be generated and knowledge-based distractor list will be generated”. System is manually evaluated through the following three steps:

• 1) . Evaluation of the selected sentence

• 2) . Evaluation of selected keyword

• 3) . Evaluation of selected distractor.

“Automatic question generation on the basis of the discourse connectives “, content selection and question generation are the two units of this system [14]. In content selection the relevant part of text is found to frame the question and question generation is includes sense disambiguation of the discourse connectives, recognition of the type of question and performing syntactic changes on the content. Two evaluators have been used for the check of semantic and syntactic correctness.

“Automatic Question Generation Using Software Agents for Technical Institutions” produced a system which took text files from the user as an input and the output hence produced by the system is a set of questionnaires based on bloom’s taxonomy [15]. Bloom’s taxonomy-based questions help in assessing the learning capability of the user. In Document processing tree tagger tool and stemming process is done to eliminate the human process [16]. Data characterization takes a rundown of keyword created by Data Processing and finds the Bloom's classification of those words, via looking through suitable activity verb in store which fits with the given keyword. The output of Information Classification is used as an input of Question Generation to produce questions. The procedure is a format-based methodology, which fits the chosen keywords in the question layout as per the Bloom's levels.

• III.  Methodology

Apart from Standford and NLTK in NATURAL LANGUAGE PROCESSING, there is a lot to enhance which system describes [14].

The methodology behind the Multiple-Choice Questions(MCQs) generation used in this proposed system includes:

• A.    Text Parsing

Either user would provide text in the form of PFD files or simply give any interesting topic, System has to parse the text especially when it comes from Wikipedia. STMS uses Beautiful Soup Python library to scrap the text from URL for a given keyword. Text parsing includes removal of References[n], special characters, and links in the text that has been scrapped from Wikipedia. Text refining is the basic approach of Pre-processing of the text. An example Python code to scrape text from URL.

• 1)  from urllib2 import urlopen

• 2)  from bs4 import BeautifulSoup

• 3)    url =" https://en.wikipedia.org/wiki/NLP "

• 4)    allData = urlopen(url)

• 5)    bs = BeautifulSoup(data, "lxml")

• 6)    rawText = bs.findAll("p") #Find All

  (paragraph tags in the HTML file)

• B.    Text Summarization

In this step, Regular Expression (REGEX) plays an important role to refine text. After refining text can be summarized using SUMY Python Library which includes (LUHN, LexRank, LSA) summarization algorithms to Summarize the text and moreover system selects an average number of sentences to generate question sentences. The importance of the sentences can be analysed using Keyword Frequency Method.

• 1)   import re

• 2)   import sumy

• 3)    from       sumy.parsers.plaintext       import

PlaintextParser

• 4)    from sumy.nlp.tokenizers import Tokenizer

• 5)   from     sumy.summarizers.lex_rank     import

LexRankSummarizer

• 6)   from       sumy.summarizers.luhn       import

LuhnSummarizer

• 7)    Removal of References and extra spaces + special characters using RegularExpressions

• 8)    article_text = re.sub('\[[0-9]*\]', ' ', article_text)

• 9)    article_text = re.sub('\s+', ' ', article_text)

• 10)    Summarization

• 11)    summarizer = LexRankSummarizer()

• 12)    summary1 = summarizer(original.document, 50)

• C.    Question Formation

After Sentences selection or text summarization, the system focuses on the question formation based on NER and SRL. SLR (Semantic Role Labelling), one of the approaches of Natural Language Processing is used to make the patterns of the words in the string of sentences.

The Named Entity Recognition (NER) comes with the same methodology to find Entities like (PERSONS, LOCATIONS, DATE, and ORGANIZATION). The System creates clusters of them and then generates questions accordingly.

• D.    Wh- Question Approach

In this approach, the system evaluates the sentences according to the NES. STMS classify the sentences like NEP (For PERSON), ORG (Organization) and LOC (Location). And generate questions accordingly.

• 1)    WHO/WHOM functions for NE(Person).

• 2)    WHERE functions call for LOC (NE-Location)

• 3)    WHEN functions call for (NE-Date).

The WH- defined functions will create complete sentences accordingly.

• E.    Pos Based

Parts-Of-Speech based questions are handled in an informal approach to ask all the relevant Proper Nouns (PNNs), which can be identified using POS tagging in python. Natural Language Tool Kit (NLTK) Python library has a major role in Natural Language Processing of the text to find all the POS tags.

• 1)    from nltk.tag import pos_tag

• 2)   stopwords                                    =

• 3)   newList = [(sent.text, sent.label_) for sent in

• 4)    qtext = word_tokenize(articcle_text)

• 5)    tags = pos_tag(ner_Text)

• 6)    print (tags)

It is needed to remove all the stop words to get the accurate Proper Nouns from a refined data.

• F.    Cloze Based

In Natural Language Processing Question Cloze based questions formation is a formal style. First of all, the system has to define the Entity to be blanked. Then RegEX can replace a certain string with that specific NE tag. The accurate answer can be placed as any of the four answers but not always at the first place. Simple for Loop can be used to change the placement of the correct answer among four of them. RegEX is used in the following code to replace the blank “_____” with certain

(NE) that System requires:

• 1)    pattern = "  "

• 2)    replacement_str = "______"

• 3)   for k, v in newList:

• 4)   if v == "PERSON":

• 5)    pattern = pattern + r'(' + k + r')|'

• 6)    pattern += r'('+ r"''" + r')'

• 7)    compiled_pattern=re.compile(pattern,re.IGNORE CASE)

• 8)    result=re.sub(compiled_pattern,replacement_str,fi nal_summary)

• 9)    each_sentence = sent_tokenize(result)

Then system tokenizes all the string of text into tokens of sentences.

• G.    Distractor Generator

One of the important steps to make an appropriate Multiple-Choice Question is to generate distractors against a correct answer. And like always Python Libraries come with a definite solution. WordNet (lexical database of English) which can be used to find Hyponyms against each correct answer and then hyponyms of each Hyponyms can help the system to find definite distractors.

e.g. Correct answer is (cat)

• a)    Cat

• b)    Rabbit

• c)    Rat

• d)    Goat

This is how hyponyms help the System STMS to complete the process of questions along with distractors.

• H.    General Work Flow Of The Algorithm

  

• I.    General Application Architecture

A three-tier architecture is used which will be client side, Web Server and DBMS.

Fig.2. A 3-Tier Architecture

• J.    Tools And Technologies

• 1.    Languages

• a)    Bootstrap: Bootstrap is used for front end development of the site which is the latest version of HTML5. This would help to create a modern and responsive site.

• b)    CSS: CSS is used for the beauty of the site which will be helpful for the modern front end of the site.

• c)    PYTHON: Python language is preferred because web scrapping for data extraction from different websites is needed and only python language has major support for web scraping in term of libraries and framework.

• d)    PHP: PHP is used for the server side of website because it would be easier to integrate with the backend process.

• e)    SQL: SQL will be used for the database creation and database handling alongside with PHP. The reason again to select SQL is that it operates well with PHP.

• 2.    Tools

• a)    Net Beans: Net Beans will be used to develop server-side code and the reason to choose Net Beans is that it has more features and provide easier development.

• b)    Anaconda: Anaconda will be used in the project to develop a web scraping code because it provides all python features at one platform.

• c)    Photoshop: Adobe Photoshop is used to develop the logo and other different graphics files for the website.

• 3.    Techniques

• a)    Data Mining: To refine the data from LEO Brain in the form of Multiple-Choice Questions(MCQs), relevant to the outline’s keywords.

• b)    Web Scrapping: Web scrapping technology will be used to extract data in the form of MultipleChoice Questions(MCQs) which will be generated by Python.

• IV.  Algorithm

• 1)    stopwords                                    =

('english') url = " "

• 2)   data = urlopen(url)

• 3)   bs = BeautifulSoup(data, "lxml")

• 4)    rawText = bs.findAll("p")

• 5)    article_text = "   "

• 6)    for p in rawText:

• 7)    article_text += p.text

• 8)    article_text = re.sub('\[[0-9]*\]', ' ', article_text)

• 9)    article_text = re.sub('\s+', ' ', article_text)

• 10)    original                                        =

PlaintextParser.from_string(article_text,Tokenizer ("english"        ))

• 11)    summarizer = LexRankSummarizer()

• 12)    summary1 = summarizer(original.document, 50)

• 13)    nlptext = nlp(final_summary)

• 14)    newList = [(sent.text, sent.label_) for sent in nlptext.ents]

• 15)    pattern = ""

• 16)    replacement_str = "______"

• 17)    for k, v in newList:

• 18)    if v == "PERSON":

• 19)    pattern = pattern + r'(' + k + r')|'

• 20)    pattern += r'('+ r"''" + r')'

• 21)    compiled_pattern      =      re.compile(pattern,

re.IGNORECASE)

• 22)    result = re.sub(compiled_pattern,replacement_str, final_summary)instances = sent_tokenize(result)

• V.  Vresults

To illustrate, system would generate Multiple Choice

Questions(MCQs)) from the given

URL of Wikipedia:

Q: The Vedic period (______BCE) was characterised by an Indo-Aryan culture: during this period the Vedas, the oldest scriptures associated with Hinduism, were composed, and this culture later became well established in the region.

(a) 1500500

(b)18th century

(c)1991

(d)1981

Q: Dispute over Jammu and Kashmir led to the First Kashmir War in______.

• (a) 1948

(b)2000

(c)1988

(d)1969

Q: The idea of Pakistan, which had received overwhelming popular support among Indian Muslims, especially those in the provinces of British India where Muslims were in a minority such as the United Provinces, was articulated in terms of an Islamic state by the Muslim League leadership, the ulama (Islamic clergy') and_______.

• (a) Jinnah

(b)Aline

(c)Theassn

(d)Joao

Q: Relations between Pakistan and Russia have greatly improved since______, and co-operation in various sectors has increased.

• (a) 1999,

  (b)2005

(c)1971

(d)1888

• Fig.3. Results Screenshot

• VI.    Discussion

In present times, unemployment is hitting its highest scores. Technology is taking over and manpower is being kicked out of the companies [17]. Regardless of whatever degrees they carry they are still being surpassed by the technology wave happening around [18]. The thought behind creating this project is to attract users to interact with knowledge domains, to learn through our easy to interact with user platforms and perform timed test sessions and to earn through this interactive learning process.

Apart from this, it will provide immense ease to students who can't find questionnaires made for their bulk of courses. And then they fail to comprehend such huge amount of uniform data and lose their scores in their examinations. This system will provide them with easy to access and easy to generate quizzes and questionnaires.

As indicated by the "Ontology Based Multiple-decision Question generation report 2015" 913 and 428 inquiries were created from the KR and Java ontologies, individually [19]. Among these are 535, 344 questions that have at any rate 3 distractors for the KR and Java ontologies, individually. Among the 913 KR questions, 50 questions were chosen haphazardly to be looked into by area specialists. Java questions were first nourished to a programmed filter that filters out any inquiry where there is a cover, i.e., a string of multiple characters that show up in both the stem and the key. This is to dispose of the odds of give some insight to the understudies to figure out the right answer without really having the expected learning to answer it effectively. These inquiries were not chosen in a simply irregular manner, rather, each inquiry has been chosen arbitrarily, at that point checked physically to ensure that it isn't repetitive, i.e., comparable regarding substance to recently chosen inquiry however extraordinary in introduction.

An inquiry is considered ''helpful'' in the event that it is evaluated as either ''valuable for what it's worth'' or ''valuable however requires minor enhancements'' by a commentator. 46 out of the 50 KR inquiries were viewed as valuable by in any event one commentator. 63 out of the 65Java inquiries have been evaluated as helpful by at any rate one commentator. A given distractor is considered ''valuable'' on the off chance that it has been useful (i.e., picked by at any rate one understudy). For the inquiries conveyed in the first round, in any event two out of three distractors were helpful. For the inquiries conveyed in the second round, at any rate one distractor was valuable with the exception of one inquiry which has been addressed accurately by all the seven understudies. We utilized Pearson's coefficient to process thing separation to demonstrate the connection between understudies' exhibition on a given inquiry and the general execution of every understudy on all inquiries. As per commentators' remarks, this is mostly because of the absence of proper articles (i.e., the, an, an).

An inquiry is too difficult for a specific gathering of understudies on the off chance that it is addressed effectively by under 30 % of the understudies and is excessively simple whenever replied by more than 90 % of the understudies. In the two cases, the inquiry should be looked into and improved. As needs be, we believe the inquiry to be difficult on the off chance that it is addressed effectively by30–60 % and simple on the off chance that it is addressed accurately by 60–90 %of the understudies. Concerning the six inquiries conveyed on paper, two inquiries were sensibly difficult and two were sensibly simple for the understudies. These four inquiries were in accordance with difficulty estimations by the Multiple-Choice Questions (MCQs) Generator apparatus. One out of the six inquiries were too difficult for the understudies. Amazingly, the device and the three commentators have evaluated this thing as simple.

At long last, one inquiry was unreasonably simple for the under studies, be that as it may, it was evaluated as difficult by the device. This is expected to having an idea in the stem. So also, for the inquiries managed on the web, one inquiry was sensibly difficult and one inquiry was sensibly simple for the understudies; just in accordance with apparatus estimations. One out of the six inquiries was unreasonably simple for the understudies (100 %correct answers). This inquiry was appraised as simple by the instrument. Once more, one inquiry was evaluated as difficult by the device yet was simple for the understudies due to having an idea in the stem. Two inquiries were not in accordance with apparatus estimations but rather were in accordance with estimations of in any event two analysts. For 51 out of the 65 Java inquiries, there has been an understanding between the apparatus and in any event one commentator. The level of understanding is a lot higher with simple inquiries achieving 100 % concurrences with at any rate one analyst. This could imply that the produced distractors for difficult questions were not conceivable enough which can be because of numerous reasons, for example, having lexical or semantic intimations. Figure 2 demonstrates the quantity of inquiries for which there is an understanding between the apparatus and at any rate one, a few analysts for the 50 KR inquiries and the 65 Java inquiries that have been checked on by the space specialists.

• VII.    Conclusion

From the comparative analysis of the question paper generated in the past out in this research, it is evident that Natural Language Processing is still a domain of huge improvements.

The system generated through the extensive research and hard work that we have employed in our project has such techniques that create questions first by scrapping the given text either in pdf or through a keyword given query. Scrapping is done through a library named BeautifulSoup which is a library of python language. Then the scrapped data is cleaned using RegEx. Cleaning basically removes any references, links or any special characters in the text. In addition, the summarization of the cleaned text is carried on using the Summy library of python. This basically summarizes the whole text to present it into a compiled form. Then this summarized text is used for Sentence Tokenization where the text is identified that where the sentences end. Using this text the Named Entity Recognizers (NER) are found. These NERs are found using Natural Language Processing (NLP) which employs the library named Natural Language ToolKit. From here the data is chunked using RegEx. This creates the questions and then distractors are generated using WordNet library.

To sum up, this system closely works on precision and generates cloze based questions.

• VIII.    Future Work

As AI is the field of precision and not accuracy. It is intended to improve the percentage of this accuracy making it near to perfect in future. Secondly this algorithm holds the system back in generating questions of subjects like chemistry maths or statistics. In future, we are looking for the solutions of this hurdle and planning to make the system smart enough to recognize the written text and generate questions without any human help.