|
def inference(self, sentence: List[str]) -> Tuple[List[int], List[float]]: |
|
probabilities = np.zeros((len(self.Q), len(sentence))) |
|
transitions = np.zeros((len(self.Q), len(sentence))) |
|
# Forward calculation |
|
for i, token in enumerate(sentence): |
|
observation = 0 |
|
if token in self._reversed_V: |
|
observation = self._reversed_V[token] |
|
|
|
if i == 0: |
|
for index, _ in enumerate(self.Q): |
|
probability = self.Pi[index] * self.B[index][observation] |
|
probabilities[index][i] = probability |
|
transitions[index][i] = -1 |
|
else: |
|
for index, _ in enumerate(self.Q): |
|
candidates = [] |
|
previous = probabilities[:, i - 1] |
|
for previous_index, previous_probability in enumerate(previous): |
|
probability = ( |
|
previous_probability |
|
* self.A[previous_index][index] |
|
* self.B[index][observation] |
|
) |
|
candidates.append(probability) |
|
|
|
probabilities[index, i] = np.max(candidates) |
|
transitions[index, i] = np.argmax(candidates) |
|
|
|
# Choose best probability path |
|
best_path = [] # type: List[int] |
|
best_probabilities = [] # type: List[float] |
|
for i in reversed(range(len(sentence))): |
|
probability = np.max(probabilities[:, i]) |
|
max_index = int(np.argmax(probabilities[:, i])) |
|
best_probabilities.append(probability) |
|
if i == len(sentence) - 1: |
|
# Last state |
|
best_path.append(max_index) |
|
if i > 0: |
|
# Previous state |
|
previous = int(transitions[max_index, i]) |
|
best_path.append(previous) |
|
|
|
best_path = best_path[::-1] # reverse the order |
|
best_probabilities = best_probabilities[::-1] # reverse the order |
|
return (best_path, best_probabilities) |
