Description:Field of Invention
A growing field of study and practical application, natural language processing (NLP) investigates the ways in which computers can comprehend and process spoken or written language in order to accomplish meaningful tasks. To program computers to comprehend and manipulate natural languages to carry out certain tasks, Professionals in natural language processing (NLP) try to learn as much as they can about how humans process and utilise language. NLP is known to be a combination of Natural Language Understanding and Natural Language Generation. Natural Language Understanding (NLU) is a field of study that deals with machine reading comprehension. It is a huge set of computer-based applications that focusses on understanding simple tasks related to natural languages such as short commands issued to robots. Analysis of full comprehensions or newspaper articles also lies under this category. Natural Language Generation (NLG) started out as the basic application where human languages were generated. But then it evolved into a software process that aimed at transforming structured data in natural languages. These natural languages can be understood both by humans and machines.
Background of the Invention
The concept of NLP evolved in the 1950s with the advent of Artificial Intelligence (AI) and linguistic studies. In the early years, NLP started out as a field of study that focused on text mining and information retrieval. But as the years progressed, it originated as a domain for numerous AI-based applications like machine translation, query expansion, robotic commands identification, etc. The foundations of NLP lie in several disciplines, viz. computer and information sciences, linguistics, mathematics, electrical and electronic engineering, artificial intelligence and robotics, psychology, etc. Answer Sheet Evaluation is a useful application of NLP is the answer sheet evaluation. Answers written by students can be well evaluated automatically using NLP techniques. Disambiguation is needed here because every student writes the answer in a unique way. Each student can use different words to describe the same theory. Optical Character Recognition (OCR) is defined as the technology that converts an image into a machine-readable and editable format. Since OCR deals with image capturing hence it is also natural for it to be susceptible to a lot of errors. These errors are corrected using NLP techniques. For the substitution of the correct word in place of the erroneous word, disambiguation techniques are applied.
OCR aims at converting text on the image in a machine-readable and editable format. The input could be in the form of any image format. The image is obtained by scanning the relevant text/pages from a huge set of concerned documents. OCR has numerous applications in day-to-day life. Traffic challans are generated after Automated Number Plate Recognition (ANPR) in vehicles. This is possible due to OCR. The old books and history related documents can be preserved for a longer time when they are stored digitally using OCRs. Since OCR deals with image processing hence it is also natural for it to be susceptible to a lot of errors while image capturing and conversion. These errors are corrected using NLP techniques. For the substitution of the correct word in place of the erroneous word, disambiguation techniques are applied. For instance, if “I am going to the Aank of the river” is written, then it is easy to detect that “Aank” is the erroneous word. But for replacing it with the correct word, the WSD algorithm might be needed. Then we can select the most appropriate word from the set of available options like Bank, Tank, etc. OCR development has been considered in the literature, but no work has been done for error detection and correction in it (US11893345B2).
The training and recognition phase in OCR can be performed using fuzzy logic. Since the characters have to be recognized precisely, the vagueness in character identification has to be eliminated, for which fuzzy logic serves the purpose. The membership values can be assigned according to the degree of resemblance of a particular alphabet with the existing character set members. Neural networks are gaining popularity among experts in various domains to optimize the given problem. In OCR developments, neural networks are deployed to instantiate the next pixel value in terms of knowledge derived from previous ones. This technique is particularly useful in scenarios where noisy text data is present. Features of a particular character or an alphabet help in defining its resemblance score. These features could be anything ranging from the length of the character to its width. Structural analysis can be performed once the feature selection process is successfully completed. Pattern recognition has always been a promising machine learning approach to solve numerous problems. In OCRs too, pattern recognition algorithms can be deployed to recognize the characters. Examples of such methods include adaptive clustering techniques. The characters are continuously tested if they possess any type of abnormal width. If they have it, it might be treated as a result of faulty identification. Hence, this technique can also help in performing error detection. Also, numerical factors such as the correlation are calculated to find out the relevance of identification and degree of resemblance of a particular character under concern. The results of adaptive clustering are known to be better as compared to the classic template matching methods, but they prove to be faulty when tested on a very large dataset. This states the concern in developing a scalable algorithm for OCRs (US10366309B2).
Short answer evaluation is an extensively popular topic among experts all over the globe. Brief answers written by students can be well evaluated automatically. But since a student can use different words for describing the “ideal answer” hence disambiguation of all content words is needed. Several approaches have been adopted by the professionals to find an optimal solution to this problem but the search for an efficient method is still going on. But it is worth noting here that all the papers have a common setback that they offer high response time. For quick and real-time applications, this response time must be low i.e., quick evaluation is needed. Also, WordNet has never been used in this context. WordNet is an online lexical database for the English language which groups words into sets of synonyms called Synsets and provides short definitions and usage examples.
Summary of the Invention
Optical Character Recognition (OCR) is the technology that converts an image into a machine-readable and editable format. OCR is discussed in literature due to its vast areas of application ranging from Handwriting Recognition to the digitization of historical data. But OCRs are susceptible to errors due to faulty recognition. It can be observed that no significant work has been done in the literature related to the error detection and correction of OCR text. In this scenario, there is a strong need to develop a methodology that can upgrade the performance of OCR using error detection & correction. Automatic answer sheet evaluation remains a grave issue in NLP since this must be a realtime task that requires the process to be impartial. This process requires disambiguation as well since every student can describe the answer in his own way using different words. After carefully evaluating the literature, it has been observed that the state-of-art methods yield high response time. But since the purpose of automation here is to decrease the time taken for evaluation hence this gap must be bridged. In this chapter, two algorithms are proposed for achieving the same. Both the algorithms use weighted WordNet graphs which have Synsets as the nodes and semantic relations as the edges.
Brief Description of Drawings
Figure 1: WordNet graph for Ideal Answer.
Detailed Description of the Invention
Optical Character Recognition (OCR) is the technology that converts an image into a machine-readable and editable format. This input could be in the form of any image format like .jpg, .png, etc. Several tools are not available online to serve the basic purpose of Optical character recognition i.e., image scanning and conversion into a human-readable format. The process of OCR can be summarized as follows: The text to be converted into machine editable form is taken as an input. If the input text is degraded in terms of quality i.e., noisy data is present in it for instance low quality photocopied documents, then the results of OCR may be erroneous. Hence it is suggested that for any kind of processing, especially for a huge database of documents, high-quality data must be taken as an input. Good quality data does not ensure error-free OCR text but the presence of noise definitely affects the overall accuracy of an OCR system. The image of the input is scanned and stored in the form of a .jpeg/ .png file. Any other image file format can also be used but it should match with the overall compatibility of the system. In image preprocessing, all the unwanted distortions are removed from the sample image so that a better quality scanned version of the image can be stored in the database. Image segmentation is highly significant as it affects the output of the consequent steps that follow it. The goal of this step is to obtain the image in the form of smaller units or pixels so that a better interpretation for the same can be achieved. The output of step 3 is fed into this step as an input. The output of previous step is fed into an output file and is displayed to the user. If some processing is to be done on this data, then it is stored as a text file. This processing can be in the form of text modification (editing) or even translation. The sequential flow of execution from step 1 to step 6 forms the base for the working of an OCR. Several online applications are available for performing OCR, such as FreeOCR, Capture2Text, etc. However, it is to be noted that during the process of OCR, some errors might occur. For instance, “sunny” could be read as “sunng”. These kinds of errors reduce the overall accuracy of an OCR and lower its performance. Word Errors occur when the word is misread as some other word and that misread word is a legitimate word according to standard dictionaries. For instance, if “king” is misread as “sing” then this becomes a word error. These errors are harder to detect and rectify. Non-Word Errors errors occur when the word is misread as some other word and that misread word is not a legitimate word according to standard dictionaries. For instance, if “king” is read as “uing” then this is a case of non-word errors. These errors commonly occur during OCR conversions and are easier to detect, although their rectification is not easy. Huge efforts have been made to create good quality OCRs, but no work has been done to create a context-based OCR that can incorporate a robust error detection and correction technique in it for better performance.
The correct “context” of a given word is used for error detection and consequent correction. This “context” establishment is done using WordNet graphs as shown in figure 1. The process initiates by converting the contents of any document under consideration into a text file. The next step is to find out if any error occurs in the text file for the sample text. Initially, the text is tokenized, and each token is compared to the existing words in an existing standard dictionary to find out if any non-word error is there in the sample text or not. WordNet is used to accomplish this task. Once the erroneous word is detected, the next step is to find a set of candidate words for it, which could be the original word as on the document. In order to generate this set, the 1-Edit Distance method is used. After figuring out the candidate words, a WordNet graph is drawn for the same using semantic relations hypernym, hyponym, holonym, and meronym which establish context and consequent perform word sense disambiguation. The WordNet graphs are generated using a depth-first search algorithm. The significance of every node in the Fuzzy WordNet graph is calculated using fuzzy centrality measures like Fuzzy Degree Centrality, Fuzzy Betweenness Centrality, Fuzzy Closeness Centrality, Fuzzy PageRank Centrality. The weight of the edges for calculating the above centrality measures are as follows: Hypernym=0.5, Hyponym=0.5, Holonym=0.5, Meronym=0.5. The node with the highest value of all the centrality values is taken for disambiguation. This disambiguated word is the original word that was meant to be in the document but got incorrectly recognized. Hence error detection and correction are performed accurately.
Using WordNet graphs and centrality indicators, the proposed method extracts key phrases for evaluating short answer sheets. The procedure begins with the scanning of each student's response form. The responses of the students are used to develop a form that can be read by machines. The student's response is tagged using the part-of-speech. In order to build a fuzzy WordNet graph, the content words are selected using a depth-first search approach, making advantage of semantic relations like hypernym, hyponym, holonym, and meronym. Because of the significance of these semantic relationships, IS-A relations will be assigned more weight than HAS-A relations in this edge weight assignment. Since the teacher is also involved in activities allied to the field of teaching, this leaves him/her very less time to completely dedicate themselves towards answer sheet evaluations. But this task is crucial and needs proper focus from the evaluator. Also, this needs to be an impartial task. In recent years, technology has evolved a lot regarding automation of answer sheet evaluation. But due to the diverse nature of human languages, vocabulary, and handwriting of the students, high levels of accuracy have not been achieved. Earlier attempts in this field have either been not so accurate or they have been extensively time-consuming. Through this proposed approach, we aim at combining the concepts of natural language processing like context identification and text similarity into question answering, to facilitate the process of short answer-based script evaluation in an accurate and time-efficient manner. One of the approaches that form the state-of-art for Short Answer Evaluation (SAE) is the Superlative model. This model begins the process by taking input from the user in the form of text or paragraphs. Then it creates a word cloud for the text under consideration. Word clouds are sometimes referred to as Tag Clouds. They are the cluster of words in which each word is depicted in different sizes. One rule establishes here for word clouds i.e., the bigger/bolder a word appears in the cluster, the more often that word is mentioned in that text. This also means that a particular word is more significant as compared to the rest. In the Superlative model, the word cloud is created for the student’s answer. If the relevant key terms, concepts, and notions are appearing bigger and bolder in that word cloud, then that answer is considered closer to the ideal answer. But it can be clearly observed that this approach suffers from a setback. The manual evaluation of the word cloud needs time. This means that the time complexity of this method is high. This defeats the sole purpose of automated short answer evaluation i.e. reduction in time for answer sheet checking. Hence the need for a fully automated system is there which is time-efficient as well.
Evaluating short answer sheets using the suggested method involves extracting important terms from WordNet graphs and centrality indicators. When a student's response sheet is scanned, the process starts. A machine-readable form is created from the student's responses. The part-of-speech is used to tag the student's response. Utilising semantic relations such as hypernym, hyponym, holonym, and meronym, a depth-first search technique is used to choose the content words in order to construct a fuzzy WordNet graph. As a result of this edge weight assignment, IS-A relations will be given higher weight than HAS-A relations, which reflects the relevance of these semantic links. , Claims:The scope of the invention is defined by the following claims:

Claim:
1. A System/Method for Answer sheet evaluation using OCR and Wordnet Graphs with machine learning techniques comprising the steps of
a) The words were extracted from the given set of answer sheets and pre-processed by removing noisy words.
b) The graphs are generated from the preprocessed words to explore the similar context words.
c) The answer sheets will be evaluated based on the similarity of the context based words generated from the answer sheet with the original answer.
2. According to claim 1, the fuzzy Optical Character Recognition is used to extract the words from the answer sheets.
3. According to claim 1, the wordnet graphs are used to explore the similar context words of the generated words and to evaluate the Answer sheets wordnet graph based similarity measures were used.