mousai

This project uses texts from project gutenberg and the Brown NLP Corpus to generate a series of lines that should be semantically correct. A parallel version will follow once the HMMTagger has been fixed and the program has been fully tested.

// Scan Books or other texts that do not have a lot of symbols to build vocabulary
→ Scan input files: 
	→ Extract sentences using end-of-line markers [? . !]
	→ Clean sentences by removing extra spaces among other unnecessary symbols
	→ Build model of sentence lengths
→ Build corpus vocabulary by splitting sentences and calculating word frequency

// Initialize Markov Chain based on in input texts
→ Initialize bigram and unigram prefix dictionary { bi/unigram : Array Of Prefixes }
→ For every sentence:
	→ Update the dictionary with prefixes and suffixes (duplicates allowed)


// Initialize + Train POS Tagging VMM Using Brown Corpus
→ Scan Brown Corpus input files:
	→ Extract word-POS pairs
	→ Extract POS patterns using end-of-sentence markers as delimiters

// Genetic Algorithm (GA)
→ Initialize ⇒ Build population of N individuals, with each individual having:
	→ A Sentence generated using the Markov Chain trained by the input texts/books
	→ A POS Tag Sequence for the sentence [Initialized with an empty ArrayList]
	→ A Fitness value of the sentence [Initialized as 0]

→ Evaluation ⇒ For each individual in the population:
→ Estimate the individual/sentence’s POS Tag Sequence using the Viterbi 
Algorithm from the VMM
→ Calculate the individual’s fitness value based on the POS Tag frequency in the Brown Corpus. As of now there is no smoothing algorithm implemented for patterns that are not recognized. 

	→ Selection ⇒ This is where the elitist part of the algorithm is introduced:
		→ If population has K individuals with the required fitness:
			// This part hasn't been implemented because I haven't found a good enough 
			// threshold to use. Shouldn't be too hard to implement though
			→ BREAK FROM GENETIC ALGORITHM
		→ Else:
			→ Sort the Individuals in terms of fitness
			→ Pick the N / 2 fittest individuals [Fitness > 0]
			→ If individuals with Fitness > 0 are less than N / 2:
				→ Discard Individuals with fitness = 0
				→ Replace them by a new set of randomly generated individuals
				→ GOTO: Evaluation


	→ Reproduction / Crossover ⇒ Do this N / 2 times:
		→ Randomly pick 2 parents from the N / 2 individuals
		→ Duplicate them
		→ Cross the duplicates at a position determined by the VMM training set
		→ Add the duplicates (children) to the population

	→ Mutation ⇒ For every individual in the population:
		→ Generate a random double
		→ If the double is less than the mutation rate
			→ Replace a random word with another word with the same POS tag
		→ GOTO: Evaluation