Description

Term Project Proposal

Group Nick (= Name of the Program)
RIDDLER
Members
(S1) Zeynep Cankara (CONTACT)
(S1) Ege Şahin
(S1) Aybars Altınışık
(S1) Göktuğ Gürbüztürk
(S1) Ahmet Feyzi Halaç

Description
In this project, we’ll write a program called RIDDLER. We’ll use the Python programming language. Our program will input a 5×5 New York Times mini-puzzle (cf. https://www.nytimes.com/crosswords/game/mini) and its solution with an automated browser interaction library. We are planning to use Stanford’s Natural Language Processing library to find the possible words corresponding to all clues one by one then together with various heuristics such as semantics and orthographic patterns, our model will decide on the correct one and fill the puzzle. In the end, we will compare our solution with the original solution to evaluate the performance of our algorithm.
Literature
The structure of Crossword puzzles relies heavily on the two aspects of the lexical memory which are clues represented as semantic information and orthographic patterns to constrain the solution space and provide hints for potential answers. The authors developed a computational model for crossword solving via adopting Biologically-Inspired Artificial Intelligence approach which models the human expert crossword play from lexical access routes and solution strategies. The authors demonstrated semantic retrieval fluency is essential for solving crossword puzzles like real life experts. Conventional Artificial Intelligence search algorithms use a trial-and-error approach when generating the correct partial solution whereas the authors algorithms mimics real-life crossword experts when making a response.
Thanasuan, Kejkaew, and Shane T. Mueller. Crossword Expertise as Recognitional Decision Making: An Artificial Intelligence Approach. 26 Aug. 2014, www.frontiersin.org/articles/10.3389/fpsyg.2014.01018/full.

Crossword expertise as recognitional decision making: an artificial intelligence approach
Kejkaew Thanasuan and Shane T. Mueller*

Edited by:
UK
Reviewed by:
Warwick, UK
*Correspondence:
Shane T. Mueller, Department of
Cognitive and Learning Sciences,
1400 Townsend Drive, Houghton, MI 49931, USA e-mail: shanem@mtu.edu
Keywords: crossword puzzles, recognitional decision making, AI, expertise, lexical memory search

1. INTRODUCTION
Crossword puzzles were first introduced in 1913, and have become both a popular pastime, mental training aid, and a domain of study for psychological researchers (e.g., Nickerson, 2011), who have long acknowledged the role of memory access in puzzle solving. Previously, Mueller and Thanasuan (2014) we proposed a model of the basic memory search processes involved in solving individual crossword clues, and suggest that the joint access and constraint provided by cues in crossword puzzles make it similar to expert decision making in many domains.
For many of the same reasons that make them engaging puzzles for humans, crossword puzzles also pose an interesting problem for Artificial Intelligence (AI) systems, as solving them requires using many of the fundamental aspects of modern AI: search, heuristics, constraint satisfaction, knowledge representation, optimization, and data mining. Because crossword solving requires searching simultaneously within two distinct spaces (i.e., semantic and orthographic), and easily permits backtracking and recursion, it is also a useful problem for learning and teaching AI (e.g., Ginsberg et al., 1990; Harris et al., 1993; Shazeer et al., 1999; Littman et al., 2002). “Dr. Fill” (Ginsberg, 2011) is currently the best-known and most advanced AI crossword solver, and it typically performs perfectly on nearly all “straight” puzzles, only making mistakes on puzzles with complex or unusual themes or letter arrangements (Lohr, 2012). For example, when competing at the 2012 American Crossword Puzzle Tournament (ACPT), Dr. Fill failed on a puzzle in which many of the answers were required
www.frontiersin.org
to be filled in backward, a twist that also challenged many human solvers. Dr. Fill finished the 2012 ACPT 141st of approximately 600 contestants and improved to 92nd place in 2013, and 67th place in 2014. The improvement over time is related not only to broader knowledge corpora being used, but also the incorporation of more rules for handling tricky puzzle themes, which often include puns, rebuses (i.e., letter substitutions), and other wordplay devices.
Although AI crossword solvers can complete many puzzles almost perfectly, these systems tend not to be based on human strategies or known human memory structure. In this paper, we