A Blazor Server app for finding good 'Codewords' for the game Codenames
C# is relatively new to me, and Blazor is totally new - this was an interesting project I took on to learn more about NLP, dotnet and C#.
Currently comes with a tiny vector embedding model - I suggest using a larger one from vectors.nlpl.eu.
The rules of the game are for a codemaster to provide a number N and a 'Codeword' that relates to N words on the board, the rest of the team must try to guess these related words, while avoiding unrelated neutral, opposing, and assassin words.
We're using a Word2vec model - a method of embedding a text corpus in a vector space, allowing us to perform similarity searches, generate representaitons of multiple words, and do some other cool things.
The general algorithm for generating codewords is as follows:
-
Create a board state, specify the number of related words desired, specify the team colour - send this to a controller.
-
Generate all permutations of words from your team's remaining cards.
-
Add the words in each permutation together to form a vector representation.
-
Find the nearest words to those representations (these are our candidate codewords).
-
For each candidate, compute its initial score from its 'distance' to the representation in step 3 multiplied by a positive coefficient.
-
From the initial score, deduct points for the cosine similarity between 'candidate' and neutral, opposing, and assassin card(s).
-
Select a shortlist of good codewords for each permutation.
-
Select a shortlist of best permutations based on their best codewords.
How to implement a fuzzy searching service for C#.
How blazor server works
How to listen to state changes between multiple components - syncrhonising two separate components
Inverse problems: assessing codewords & figuring out the words from a codeword
Better namespace segregation
More closing following SOLID principles