Fix dep-tags

dvc.yaml

@@ -379,12 +379,12 @@ stages:
             -   dataset: pl_common_voice
                 asr: ajn
 
-            -   dataset: pl_voicelab_cbiz
-                asr: google
-            -   dataset: pl_voicelab_cbiz
-                asr: ajn
-            -   dataset: pl_voicelab_cbiz
-                asr: techmo
+#            -   dataset: pl_voicelab_cbiz
+#                asr: google
+#            -   dataset: pl_voicelab_cbiz
+#                asr: ajn
+#            -   dataset: pl_voicelab_cbiz
+#                asr: techmo
 
             -   dataset: pl_google_fleurs
                 asr: google

experiment/sentence_wer_processor/spacy_pos_sentence_dep_tag_processor.py

@@ -25,4 +25,4 @@ class SpacyDepTagSentenceWerProcessor(SentenceWerProcessor):
 
     def process_text(self, text: str) -> List[str]:
         document = self._nlp(text)
-        return [token.pos_ for token in document]
+        return [token.dep_ for token in document]

experiment_data/pipeline/pl_common_voice/.gitignore

@@ -40,3 +40,7 @@
 /wav2vec2__spacy_dep_tag_metrics
 /ajn__spacy_pos_alignment
 /ajn__spacy_pos_metrics
+/google__spacy_dep_tag_alignment
+/google__spacy_dep_tag_metrics
+/techmo__spacy_dep_tag_alignment
+/techmo__spacy_dep_tag_metrics

experiment_data/pipeline/pl_google_fleurs/.gitignore

@@ -57,3 +57,5 @@
 /google__word_wer_embeddings_alignment
 /google__spacy_dep_tag_alignment
 /google__spacy_dep_tag_metrics
+/wav2vec2__spacy_dep_tag_alignment
+/wav2vec2__spacy_dep_tag_metrics

experiment_data/pipeline/pl_luna/.gitignore

@@ -57,3 +57,5 @@
 /google__spacy_dep_tag_metrics
 /techmo__spacy_dep_tag_alignment
 /techmo__spacy_dep_tag_metrics
+/wav2vec2__spacy_dep_tag_alignment
+/wav2vec2__spacy_dep_tag_metrics

experiment_data/pipeline/pl_minds14/.gitignore

@@ -49,3 +49,11 @@
 /ajn__word_wer_classic_alignment
 /ajn__word_wer_embeddings_metrics
 /ajn__word_wer_embeddings_alignment
+/wav2vec2__spacy_dep_tag_alignment
+/wav2vec2__spacy_dep_tag_metrics
+/google__spacy_dep_tag_alignment
+/google__spacy_dep_tag_metrics
+/techmo__spacy_dep_tag_alignment
+/techmo__spacy_dep_tag_metrics
+/ajn__spacy_dep_tag_alignment
+/ajn__spacy_dep_tag_metrics

sziszapangma/core/alignment/distance_matrix_calculator.py

@@ -68,7 +68,12 @@ class CosineDistanceCalculator(DistanceCalculator):
             raise RuntimeError("array dimensions {} not right".format(a.ndim))
         similarity = np.dot(a, b.T) / (a_norm * b_norm)
         dist = 1.0 - similarity
-        return float(dist)
+        # return float(dist)
+        float_dist = float(dist)
+        if abs(float_dist) < 0.000001:
+            return 0.0
+        else:
+            return float_dist
 
     def calculate_distance_matrix(
         self, reference: List[Word], hypothesis: List[Word]

sziszapangma/core/wer/wer_calculator.py

 from abc import ABC
 from typing import List
 
+import numpy as np
+
 from sziszapangma.core.alignment.alignment_step import AlignmentStep
 from sziszapangma.core.alignment.alignment_util import AlignmentUtil
+from sziszapangma.core.alignment.step_type import StepType
 from sziszapangma.core.wer.span import Span
 
 
@@ -27,7 +30,8 @@ class WerCalculator(ABC):
         steps: List[AlignmentStep],
     ) -> float:
         reference_len = AlignmentUtil.get_reference_length(steps)
-        return sum([step.step_cost for step in steps]) / reference_len
+        fixed_step_costs = [step.step_cost for step in steps]
+        return sum(fixed_step_costs) / reference_len
 
     def calculate_wer(self, steps: List[AlignmentStep]) -> float:
         return self._calculate_wer(steps)

sziszapangma/integration/task/embedding_wer_metrics_task.py

+from typing import List
+
 from sziszapangma.core.alignment.alignment_embedding_calculator import AlignmentEmbeddingCalculator
 from sziszapangma.core.alignment.alignment_soft_calculator import AlignmentSoftCalculator
 from sziszapangma.core.transformer.cached_embedding_transformer import CachedEmbeddingTransformer
@@ -7,6 +9,7 @@ from sziszapangma.integration.mapper.alignment_step_mapper import AlignmentStepM
 from sziszapangma.integration.repository.experiment_repository import ExperimentRepository
 from sziszapangma.integration.task.processing_task import ProcessingTask
 from sziszapangma.integration.task.task_util import TaskUtil
+from sziszapangma.model.model import Word
 from sziszapangma.model.relation_manager import RelationManager
 
 _SOFT_WER = "soft_wer"
@@ -51,6 +54,10 @@ class EmbeddingWerMetricsTask(ProcessingTask):
             is not None
         )
 
+    @staticmethod
+    def filter_empty_words(words: List[Word]) -> List[Word]:
+        return [it for it in words if len(it['text']) > 0]
+
     def run_single_process(
         self,
         record_id: str,
@@ -60,8 +67,8 @@ class EmbeddingWerMetricsTask(ProcessingTask):
         gold_transcript = TaskUtil.get_words_from_record(relation_manager)
         asr_result = experiment_repository.get_property_for_key(record_id, self._asr_property_name)
         if gold_transcript is not None and asr_result is not None and "transcription" in asr_result:
-            gold_transcript_lower = TaskUtil.words_to_lower(gold_transcript)
-            asr_transcript_lower = TaskUtil.words_to_lower(asr_result["transcription"])
+            gold_transcript_lower = self.filter_empty_words(TaskUtil.words_to_lower(gold_transcript))
+            asr_transcript_lower = self.filter_empty_words(TaskUtil.words_to_lower(asr_result["transcription"]))
 
             soft_alignment = self._alignment_soft_calculator.calculate_alignment(
                 gold_transcript_lower, asr_transcript_lower
@@ -80,6 +87,7 @@ class EmbeddingWerMetricsTask(ProcessingTask):
                 ],
             }
             wer_results = {"soft_wer": soft_wer, "embedding_wer": embedding_wer}
+            print(wer_results)
 
             experiment_repository.update_property_for_key(
                 record_id, self._alignment_property_name, alignment_results