Please use this identifier to cite or link to this item: https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4622
Title: End-End Automated Crossword Puzzle solver using image processing, NLP and neural network
Authors: Gammedda, T.L.H.
Keywords: Crossword Puzzle Solver
OpenCV
Natural Language Processing
skip-gram, Continuous Bag of Word
Issue Date: 20-Jul-2022
Abstract: Solving newspaper crosswords helps to improve human brain function and mental health. Therefore, it has become a popular word game all over the world. It also helps decrease depression and anxiety that occurs due to social isolation and loneliness in the COVID-19 pandemic. In this game, players are very enthusiastic to find the correct answer after solving the puzzle. However, In newspaper crosswords, to check the answers, players have to wait a week or more than one week until that particular paper publishes the answers. Due to this delay, players are moving to online crosswords that provide a solution quickly. Mainly this issue occurs in Sri Lanka. Therefore, this research aimed to provide an automatic crossword solver that can motivate people to play newspaper crosswords. The importance of the proposed system is that it can solve different clue types in British-style crosswords using the image of the crossword puzzle. There is no way to check the answers using existing puzzle solvers because they contain several limitations. One of the limitations is, players should have to have the technical knowledge to use the solver. Being unable to provide accurate answers for different clue types in British-style crosswords is another biggest limitation of current solvers. The proposed system is an automated crossword puzzle solver which can be used to solve crosswords using the image of the crossword by resolving these issues. Cell extraction module, clue categorization module, candidate list generation module, filtering module, and neural network module are the main components of this implemented system. OpenCV was used to extract the cell details and used the Python-tesseract to extract the details from the clue image. Clue preprocessing was done using Natural Language Toolkit (NLTK) and categorized the preprocessed clues into clue types namely fill-in-the-blanks, synonyms, anagrams, and definition type clues. Generated the candidate answers for each clue using OneLook reverse dictionary, anagram solver, Web scraping, and Word2vec model based on the type of the clue. These candidate answers were filtered using length constraints and overlap positions of the puzzle grid. Then, a filtered candidate answer was sent to the neural network module, and identified the correct answer for each clue with the help of the trained model. Finally, the puzzle grid was filled using OpenCV. The proposed system achieved more than 90% of correct words and letters in the evaluation.
URI: https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4622
Appears in Collections:2021

Files in This Item:
File Description SizeFormat 
2018 MCS 022.pdf16.62 MBAdobe PDFView/Open


Items in UCSC Digital Library are protected by copyright, with all rights reserved, unless otherwise indicated.