NER pipeline for detecting minor children and author gender indications in product reviews, with an optional medical-entity extension. Multiple model backends are provided so accuracy, speed, and zero-shot generalization trade-offs can be compared directly.

Adds three biomedical labels from clinical case reports:

CSV column format (minor_col, gender_col): semicolon-separated span phrases that appear verbatim in ori_review. Empty = no entity of that type in the row.

Eight backends are implemented. All scripts default to combined_upload.csv (DeBERTa) or mydata.csv (others); pass --csv to override.

Install only the backend you intend to run:

pip install -r requirements-spacy.txt      # spaCy
pip install -r requirements-spanmarker.txt # SpanMarker
pip install -r requirements-gliner.txt     # GLiNER / GLiNER2
pip install -r requirements-flant5.txt     # FLAN-T5
pip install -r requirements-deberta.txt    # DeBERTa
pip install -r requirements-llm.txt        # Qwen LLM (LoRA)
pip install -r requirements-setfit.txt     # SetFit

Token-classification NER using a spaCy pipeline (BIO tagging). Fast inference; base-model general NER is not retained after fine-tuning.

python minor_gender_only/train_spacy_csv.py [options]

Outputs: best/, final/, checkpoint-epoch-N/

Span-level BERT classifier using BIO supervision. Best accuracy on fixed label sets; saves best/ by eval_overall_f1 (SpanMarker's built-in seqeval metric).

python minor_gender_only/train_spanmarker_csv.py [options]

Outputs: best/, final/

Fine-tunes a GLiNER span model. Retains zero-shot generalization for unseen entity types after fine-tuning.

python minor_gender_only/train_gliner_csv.py [options]

Outputs: best/, final/

Fine-tunes GLiNER2 with LoRA adapters. Uses natural-language label descriptions instead of bare label names, which improves implicit entity detection.

python minor_gender_only/train_gliner2_csv.py [options]

Outputs: best/, final/

Seq2seq instruction-tuned model that generates a JSON array of entities. The descriptive label names in the prompt give the model a semantic head-start.

python minor_gender_only/train_flant5_csv.py [options]

Outputs: best/, final/

Token-classification NER using DeBERTa-v3-base. Uses character-offset-based label alignment with SentencePiece tokenization. Defaults to combined_upload.csv.

python minor_gender_only/train_deberta_csv.py [options]

Outputs: best/, final/

Note: DeBERTa-v3's SentencePiece tokenizer includes the preceding space in subword offsets (e.g. ▁baby → offset (2, 7) rather than (3, 7)). The eval callback accounts for this by comparing stripped span text rather than raw character indices.

Fine-tunes Qwen3.5-9B with 4-bit quantization and LoRA. The model generates a JSON array of {"text": ..., "label": ...} objects. Slowest to train; best at implicit/contextual entity detection.

python minor_gender_only/train_llm_csv.py [options]

Outputs: best/, final/, checkpoint-epoch-N/

Contrastive sentence-embedding classifier over candidate n-gram spans. Trains a span-level classifier: each candidate span is embedded and classified as NonfictionalChildRelated, AuthorGenderIndication, or O. No generative decoding; very fast inference.

python minor_gender_only/train_setfit_csv.py [options]

Outputs: best/, final/

Evaluates any combination of trained models on a CSV file. Uses corpus-level and row-average soft F1 (exact match + token-Jaccard partial credit).

python minor_gender_only/evaluate_csv.py \
  --csv combined_upload.csv \
  --spacy minor_gender_only/spacy_model_combined_v2/best \
  --gliner gliner_finetuned_csv/best \
  --spanmarker spanmarker_finetuned_csv/best \
  --keyword-baseline \
  --row-output eval_results.csv

Evaluates the spaCy model against one of the synthetic_*_edge_cases.txt files, which test known hard cases (pet vs. child language, fictional vs. real children, age edge cases, etc.).

python minor_gender_only/eval_edge_cases_spacy.py \
  --model minor_gender_only/spacy_model_combined_v2/best \
  --edge-cases synthetic_700_edge_cases.txt \
  --errors \
  --row-output edge_results.csv

Zero-shot baseline: uses BART-large-MNLI to detect whether a sentence contains a minor/gender entity, then falls back to keyword extraction for the span. No fine-tuning required.

python minor_gender_only/zero_shot_bart_mnli_csv.py --csv combined_upload.csv

Trains on MACCROBAT clinical case reports + Corona2.json + mydata.csv in a single pass, adding MedicalCondition, ClinicalProcedure, and ClinicalEvent labels.

Required data:

• 9764942/MACCROBAT2018/ — BratStandoff .txt + .ann files
• Corona2.json — place next to the scripts

python with_medical/train_spacy.py      # spaCy
python with_medical/train_gliner.py     # GLiNER
python with_medical/train_spanmarker.py # SpanMarker
python with_medical/train_llm.py        # Qwen LLM

See the argument tables in the old README section or run python <script> --help. All scripts share the same --data-dir, --corona, --csv, and --minor-oversample arguments as their minor_gender_only/ counterparts.

Utility scripts for preparing and auditing the annotation data.

remove_minor_pronouns.py (project root) strips reviewer pronouns from gender/minor columns to reduce noise from ambiguous pronoun labels.

The synthetic_*_edge_cases.txt files are Python dicts (loadable with ast.literal_eval) covering known ambiguity categories: