How soheilaansari/persian-culture-aware-translation works
The repository releases three research capabilities: a text-only Persian→English translation baseline, its audio-conditioned extension, and a culture-specific-item (CSI) tagging and evaluation pipeline. The principal user-facing workflow is supplying English hypotheses and CSI-tagged test data plus lexicons to compute CSI-Recall. The documented stages connect corpus preparation, model training, annotation, alignment, lexicon construction, and evaluation. The CSI pipeline links are documented in the README; sampled source excerpts provide additional implementation evidence, but unsampled internals are not assumed.
Main entry points: annotate_corpus.py Corpus Annotation · Masnavi coprus Masnavi Pairs · finetune.py Parallel Fine-tuning · csi_recall_metric.py CSI-Recall
Components 19
Text Baseline
- Domain Pretraining
src/text_baseline/domain_pretrain.py - Parallel Fine-tuning
src/text_baseline/finetune.py - Translation Metrics
src/text_baseline/evaluate.py
Multimodal Translation
- Training Config
src/multi-modal/config.py - Text Audio Loader
src/multi-modal/data_loader.py - Fusion Model
src/multi-modal/model.py - Multimodal Training
src/multi-modal/train.py
CSI Evaluation
- Gold CSI Spans
- BIO Dataset Prep
src/csi_evaluation/prepare_bio_dataset.py - CSI Tagger Training
src/csi_evaluation/train_tagger.py - Corpus Annotation
src/csi_evaluation/annotate_corpus.py - Alignment Preparation
src/csi_evaluation/prepare_alignment_input.py - Awesome-align
- CSI Lexicon
src/csi_evaluation/extract_lexicon.py - CSI-Recall
src/csi_evaluation/csi_recall_metric.py - Recall Results
Research Data
- Persian Poetry
- Masnavi Pairs
data/Masnavi coprus
Other components
- English Hypotheses
Connections 22
- Persian Poetry → Domain Pretraining: trains on
README.md - Domain Pretraining → Parallel Fine-tuning: provides checkpoint
README.md - Masnavi Pairs → Parallel Fine-tuning: fine-tunes on
README.md - Parallel Fine-tuning → Translation Metrics: evaluates model
README.md - Masnavi Pairs → Translation Metrics: test split
README.md - Training Config → Multimodal Training: configures
src/multi-modal/train.py - Text Audio Loader → Multimodal Training: supplies batches
src/multi-modal/train.py - Parallel Fine-tuning → Fusion Model: initializes text encoder
README.md - Fusion Model → Multimodal Training: trained by
README.md - Masnavi Pairs → Text Audio Loader: aligned text rows
README.md - Gold CSI Spans → BIO Dataset Prep: converts spans
README.md - BIO Dataset Prep → CSI Tagger Training: trains from
README.md - CSI Tagger Training → Corpus Annotation: provides tagger
README.md - Masnavi Pairs → Corpus Annotation: annotates corpus
README.md - Corpus Annotation → Alignment Preparation: prepares tagged pairs
README.md - Alignment Preparation → Awesome-align: provides alignment input
README.md - Awesome-align → CSI Lexicon: provides alignments
README.md - Corpus Annotation → CSI Lexicon: provides tagged pairs
README.md - English Hypotheses → CSI-Recall: scores translations
README.md - CSI Lexicon → CSI-Recall: judges CSI spans
README.md - Corpus Annotation → CSI-Recall: supplies tagged test data
README.md - CSI-Recall → Recall Results: reports scores
README.md