2.0.8 -- different generation algorithm

lib/twitter_ebooks.rb

@@ -16,5 +16,6 @@ end
 require 'twitter_ebooks/nlp'
 require 'twitter_ebooks/archiver'
 require 'twitter_ebooks/markov'
+require 'twitter_ebooks/suffix'
 require 'twitter_ebooks/model'
 require 'twitter_ebooks/bot'

lib/twitter_ebooks/markov.rb

@@ -54,9 +54,10 @@ module Ebooks
 
     def chain(tokens)
       if tokens.length == 1
-        matches = @unigrams[tokens[0]]
+        matches = @unigrams[tokens[-1]]
       else
         matches = @bigrams[tokens[-2]][tokens[-1]]
+        matches = @unigrams[tokens[-1]] if matches.length < 2
       end
 
       if matches.empty?

lib/twitter_ebooks/model.rb

@@ -7,7 +7,7 @@ require 'digest/md5'
 
 module Ebooks
   class Model
-    attr_accessor :hash, :sentences, :markov, :keywords
+    attr_accessor :hash, :sentences, :generator, :keywords
 
     def self.consume(txtpath)
       Model.new.consume(txtpath)
@@ -67,16 +67,29 @@ module Ebooks
       NLP.htmlentities.decode tweet
     end
 
-    def markov_statement(limit=140, markov=nil)
-      markov ||= MarkovModel.build(@sentences)
+    def valid_tweet?(tokens, limit)
+      tweet = NLP.reconstruct(tokens)
+      tweet.length <= limit && !NLP.unmatched_enclosers?(tweet)
+    end
+
+    def make_statement(limit=140, generator=nil)
+      responding = !generator.nil?
+      generator = SuffixGenerator.build(@sentences)
       tweet = ""
 
-      while (tweet = markov.generate) do
-        next if tweet.length > limit
-        next if NLP.unmatched_enclosers?(tweet)
-        break if tweet.length > limit*0.4 || rand > 0.8
+      while (tokens = generator.generate(3, :bigrams)) do
+        next if tokens.length <= 3 && !responding
+        break if valid_tweet?(tokens, limit)
       end
 
+      if @sentences.include?(tokens) && tokens.length > 3 # We made a verbatim tweet by accident
+        while (tokens = generator.generate(3, :unigrams)) do
+          break if valid_tweet?(tokens, limit) && !@sentences.include?(tokens)
+        end
+      end
+
+      tweet = NLP.reconstruct(tokens)
+
       fix tweet
     end
 
@@ -101,19 +114,19 @@ module Ebooks
     end
 
     # Generates a response by looking for related sentences
-    # in the corpus and building a smaller markov model from these
-    def markov_response(input, limit=140)
+    # in the corpus and building a smaller generator from these
+    def make_response(input, limit=140)
       # First try 
       relevant, slightly_relevant = relevant_sentences(input)
 
       if relevant.length >= 3
-        markov = MarkovModel.new.consume(relevant)
-        markov_statement(limit, markov)
-      elsif slightly_relevant.length > 5
-        markov = MarkovModel.new.consume(slightly_relevant)
-        markov_statement(limit, markov)
+        generator = SuffixGenerator.build(relevant)
+        make_statement(limit, generator)
+      elsif slightly_relevant.length >= 5
+        generator = SuffixGenerator.build(slightly_relevant)
+        make_statement(limit, generator)
       else
-        markov_statement(limit)
+        make_statement(limit)
       end
     end
   end

lib/twitter_ebooks/nlp.rb

@@ -61,7 +61,7 @@ module Ebooks
     # As above, this is ad hoc because tokenization libraries
     # do not behave well wrt. things like emoticons and timestamps
     def self.tokenize(sentence)
-      regex = /\s+|(?<=[#{PUNCTUATION}])(?=[a-zA-Z])|(?<=[a-zA-Z])(?=[#{PUNCTUATION}]+)/
+      regex = /\s+|(?<=[#{PUNCTUATION}]\s)(?=[a-zA-Z])|(?<=[a-zA-Z])(?=[#{PUNCTUATION}]+\s)/
       sentence.split(regex)
     end
 
@@ -150,5 +150,12 @@ module Ebooks
 
       false
     end
+
+    # Determine if a2 is a subsequence of a1
+    def self.subseq?(a1, a2)
+      a1.each_index.find do |i|
+        a1[i...i+a2.length] == a2
+      end
+    end
   end
 end

lib/twitter_ebooks/suffix.rb

@@ -0,0 +1,82 @@
+module Ebooks
+  class SuffixGenerator
+    def self.build(sentences)
+      SuffixGenerator.new(sentences)
+    end
+
+    def initialize(sentences)
+      @sentences = sentences.reject { |s| s.length < 2 }
+      @unigrams = {}
+      @bigrams = {}
+
+      @sentences.each_with_index do |tokens, i|
+        last_token = INTERIM
+        tokens.each_with_index do |token, j|
+          @unigrams[last_token] ||= []
+          @unigrams[last_token] << [i, j]
+
+          @bigrams[last_token] ||= {}
+          @bigrams[last_token][token] ||= []
+
+          if j == tokens.length-1 # Mark sentence endings
+            @unigrams[token] ||= []
+            @unigrams[token] << [i, INTERIM]
+            @bigrams[last_token][token] << [i, INTERIM]
+          else
+            @bigrams[last_token][token] << [i, j+1]
+          end
+
+          last_token = token
+        end
+      end
+
+      self
+    end
+
+    def generate(passes=5, n=:unigrams)
+      index = rand(@sentences.length)
+      tokens = @sentences[index]
+      used = [index] # Sentences we've already used
+      verbatim = [tokens] # Verbatim sentences to avoid reproducing
+
+      0.upto(passes-1) do
+        puts NLP.reconstruct(tokens)
+        varsites = {} # Map bigram start site => next token alternatives
+
+        tokens.each_with_index do |token, i|
+          next_token = tokens[i+1]
+          break if next_token.nil?
+          
+          alternatives = (n == :unigrams) ? @unigrams[next_token] : @bigrams[token][next_token]
+          alternatives.reject! { |a| a[1] == INTERIM || used.include?(a[0]) }
+          varsites[i] = alternatives unless alternatives.empty?
+        end
+
+        variant = nil
+        varsites.to_a.shuffle.each do |site|
+          start = site[0]
+
+          site[1].shuffle.each do |alt|
+            start, alt = site[0], site[1].sample
+            verbatim << @sentences[alt[0]]
+            suffix = @sentences[alt[0]][alt[1]..-1]
+            potential = tokens[0..start+1] + suffix
+            
+            unless verbatim.find { |v| NLP.subseq?(v, potential) || NLP.subseq?(potential, v) }
+              used << alt[0]
+              variant = potential
+              break
+            end
+          end
+
+          break if variant
+        end
+
+        tokens = variant if variant
+      end
+
+
+      tokens
+    end
+  end
+end

lib/twitter_ebooks/version.rb

@@ -1,3 +1,3 @@
 module Ebooks
-  VERSION = "2.0.7"
+  VERSION = "2.0.8"
 end