Function get_bigram_probs

nlp_class2/markov.py:19–44

(sentences, V, start_idx, end_idx, smoothing=1)

def get_bigram_probs(sentences, V, start_idx, end_idx, smoothing=1):
  # structure of bigram probability matrix will be:
  # (last word, current word) --> probability
  # we will use add-1 smoothing
  # note: we'll always ignore this from the END token
  bigram_probs = np.ones((V, V)) * smoothing
  for sentence in sentences:
    for i in range(len(sentence)):
      
      if i == 0:
        # beginning word
        bigram_probs[start_idx, sentence[i]] += 1
      else:
        # middle word
        bigram_probs[sentence[i-1], sentence[i]] += 1

      # if we're at the final word
      # we update the bigram for last -> current
      # AND current -> END token
      if i == len(sentence) - 1:
        # final word
        bigram_probs[sentence[i], end_idx] += 1

  # normalize the counts along the rows to get probabilities
  bigram_probs /= bigram_probs.sum(axis=1, keepdims=True)
  return bigram_probs