NormForm (v0.1 — Experimental)

NormForm is a Python toolkit and command-line interface for context-aware Norwegian morphological standardization, alternative generation, and style auditing in both Bokmål (nob) and Nynorsk (nno).

Written Norwegian officially permits wide intra-norm morphological and orthographic variation—such as feminine vs. common gender nouns (boka / boken), weak preterite endings (kasta / kastet), neuter plural endings (husene / husa), and lexical spelling variants (fram / frem, sjuk / syk, no / nå). NormForm lets you:

1. Standardize text or large corpora to a consistent sub-norm style (such as Bokmål Redaksjonsnorm, Nynorsk Nasjonalbiblioteknorm, or your own custom style profile).
2. Generate [alt1|alt2] morphological alternative brackets for speech recognition (ASR) and text generation evaluation with altmetrics and altwer.
3. Audit documents and corpora for internal morphological inconsistencies or rank the exact morphological changes made across a standardized dataset.
4. Extract empirical style prompt words from any reference document using information-theoretic submodular maximization.
5. Share, fork, and version style profiles as lightweight YAML repositories on Hugging Face (1 repo = 1 style).

Table of Contents

1. What NormForm Does — and What It Does NOT Do
2. Installation
3. Quick Start (CLI & Python API)
4. Feature Guide
   • 4.1 Standardizing Text & Documents (normform standardize)
   • 4.2 High-Throughput 1-GPU Batch Corpus Standardization (normform batch)
   • 4.3 Generating [alt1|alt2] Morphological Alternatives (normform alternatives)
   • 4.4 Auditing Inconsistencies & Change Statistics (normform audit)
   • 4.5 Extracting Steering Style Words (normform extract-style)
   • 4.6 Managing & Customizing Styles (normform styles)
5. Style File Format (style.yaml)
6. Official Models & Style Repositories
7. License & Credits

1. What NormForm Does — and What It Does NOT Do

NormForm is built around a strict non-destructive lattice contract (normform.contract). The neural model is never allowed to freely rewrite your text. Instead, every sentence is parsed into a verified template of literal spans and [alt1|alt2] candidate slots (Template.align), and only validated candidate slots can change.

What NormForm DOES:

• Chooses consistently among valid intra-norm alternatives in context: e.g., selecting boken from [boka|boken], kastet from [kasta|kastet], or frem from [fram|frem] according to your chosen style profile.
• Guarantees byte-for-byte preservation of all non-alternative text: punctuation, whitespace, line breaks, numbers, dates, URLs, email addresses, proper names, and out-of-vocabulary words are never altered.
• Preserves source token casing automatically:
  • Lowercase: boka $\rightarrow$ boken
  • Titlecase (sentence-initial): Boka $\rightarrow$ Boken
  • Uppercase: BOKA $\rightarrow$ BOKEN
  • Mixed internal capitals: vebb-TV $\rightarrow$ web-TV
• Protects homographs and semantic collisions (invalid_pairs): When two surface forms can overlap in inflection tables but differ in meaning or grammatical role (e.g., verken | verket "neither" vs. "the pain/work", bad | badet, slott | slåss), the style profile's invalid_pairs guard forces NormForm to preserve the original token unchanged.
• Maintains document- and corpus-level consistency: Once a candidate group (boka | boken) is resolved—either from a reviewed style cache or by the neural style model—that exact decision is cached and applied consistently across every occurrence.

What NormForm Does NOT Do:

• Does NOT translate between Bokmål and Nynorsk: Specify --lang nob for Bokmål text and --lang nno for Nynorsk text.
• Does NOT fix general typos, grammar errors, or punctuation: Words that are not valid Norwegian forms with morphological alternatives are passed through untouched.
• Does NOT paraphrase, reorder words, or insert/delete tokens: Every output sentence is guaranteed to align 1-to-1 with the source sentence's literal template.
• Does NOT rewrite syntactic agreement across neighboring tokens: NormForm substitutes individual morphological word slots ([boka|boken]). It does not rewrite surrounding articles or adjectives outside the candidate slot (e.g., if a source phrase has ei lita bok, individual word slots are only changed where each slot's own paradigm allows it).
• Does NOT use brittle suffix-stripping heuristics on compounds: Novel compound words absent from the seed cache are either resolved in full sentential context by the neural model or preserved unchanged when --preserve-unreviewed is enabled.

2. Installation

pip install normform

# Or install with batch dataset extras (Parquet, Hugging Face Datasets, progress bars):
pip install "normform[batch]"

Requirements:

• Python $\ge 3.10$ (tested on Python 3.12 with PyTorch $\ge 2.0$ and Transformers $\ge 4.40$).
• A CUDA-capable GPU is strongly recommended for model inference and corpus batch processing (--device cuda:0 or --device auto). CPU execution is supported for small texts and --raw audits.
• No API keys required: The pinned neural models and official style repositories on Hugging Face are public and downloaded automatically on first use.

3. Quick Start (CLI & Python API)

Command Line (normform)

# Check version
normform --version

# 1. Standardize a Bokmål sentence using the official Redaksjonsnorm style:
normform standardize -t "Jenta kasta ballen og la boka på hylla." --lang nob --style redaksjonsnorm

# 2. Standardize a Nynorsk sentence using Nasjonalbiblioteknorm:
normform standardize -t "Me køyrde heim på laurdag." --lang nno --style nasjonalbiblioteknorm

# 3. Generate [alt1|alt2] morphological alternative brackets:
normform alternatives -t "Katta ligger på matta." --lang nob
# -> [Katta|Katten] ligger på [matta|matten].

Python API

from normform import Standardizer

# Initialize with a built-in style alias, Hugging Face repo ID, or local .yaml file:
std = Standardizer(
    lang="nob",
    style="redaksjonsnorm",       # or "NbAiLab/normform-bokmaal-redaksjonsnorm"
    preserve_unreviewed=False,    # True = only apply pre-reviewed cache entries
    device="auto",
)

# 1. Standardize text (preserves casing, punctuation, and layout):
text = std("Jenta kasta ballen og la boka på hylla.")
print(text)
# -> "Jenta kastet ballen og la boken på hyllen."

# 2. Generate [alt1|alt2] morphological alternative brackets:
bracketed = std.alternatives("Katta ligger på matta.")
print(bracketed)
# -> "[Katta|Katten] ligger på [matta|matten]."

4. Feature Guide

4.1 Standardizing Text & Documents (normform standardize)

Use normform standardize for interactive strings, plain-text documents (.txt), or small JSONL files (.jsonl).

# Standardize a .txt file with Redaksjonsnorm and write to output.txt:
normform standardize -i article.txt -o article_standardized.txt --lang nob --style redaksjonsnorm

# Strict reviewed-only mode (never invoke the style model on unseen compounds;
# only apply the 33,000+ human/linguist-reviewed choices in the style cache):
normform standardize -i article.txt -o article_standardized.txt \
  --lang nob --style redaksjonsnorm --preserve-unreviewed

# Standardize without a style file by supplying 5-10 steering prompt words:
normform standardize -t "Jenta hoppa høyt over gjerdet." \
  --lang nob -p "jenta,kasta,boka,hylla,fram"

# Persist choices across multiple runs in a local JSON cache and write a Markdown summary:
normform standardize -i chapter1.txt -o chapter1_norm.txt \
  --lang nob --style redaksjonsnorm \
  --cache project_choices.json --summary-output chapter1_report.md

4.2 High-Throughput 1-GPU Batch Corpus Standardization (normform batch)

For large datasets (tens of thousands to millions of records in .jsonl, .parquet, directory shards, or Hugging Face datasets), use normform batch.

How the 3-Stage Batch Engine Works

1. Background I/O Prefetch & Sentence Splitting: A background producer thread streams records from disk or Hugging Face and splits text into whitespace-preserving sentence spans so disk/network I/O overlaps with GPU compute.
2. Stage 1 — Length-Bucketed GPU Alternative Expansion: Non-blank sentences in each chunk (--chunk-size 4096) are sorted by character length to minimize padding waste, expanded in large GPU batches (--batch-size 512), and validated once with clean_direct + parse + Template.align. Over-capacity outlier sentences are automatically isolated via binary bisection and preserved unchanged.
3. Stage 2 — Greedy Set-Cover & Context-Style Resolution:
   • All candidate groups already present in your --style cache are resolved in $O(1)$ memory lookup with zero extra GPU calls.
   • If unseen candidate groups appear in the chunk (and --preserve-unreviewed is not set), NormForm runs Greedy Set-Cover (greedy_set_cover) to pick the minimum subset of sentences needed to cover all unseen candidate groups, runs adapter.style_batch() on just those sentences, and caches the resulting choices for the rest of the corpus.
4. Stage 3 — Async Streaming Rewrite & Sealed Cache Export: Sentences are rendered using the updated cache, reassembled into records, and written asynchronously to .jsonl, .parquet (Zstandard-compressed), or .txt. At completion, NormForm exports a human-editable *_sealed_cache.yaml top-sorted by actual corpus change impact (# seen: ..., changed: ..., already_correct: ..., change_rate: ...%).

Batch Examples

# 1. Standardize a large JSONL corpus on GPU 0 using Bokmål Redaksjonsnorm:
normform batch \
  -i /path/to/corpus.jsonl \
  -o /path/to/corpus_standardized.jsonl \
  --lang nob \
  --style NbAiLab/normform-bokmaal-redaksjonsnorm \
  --text-field text \
  --output-field normalized \
  --batch-size 512 \
  --chunk-size 4096 \
  --device cuda:0

# 2. Standardize a Parquet file or directory of Parquet shards in strict reviewed-only mode:
normform batch \
  -i /path/to/parquet_shards/ \
  -o /path/to/standardized.parquet \
  --lang nno \
  --style NbAiLab/normform-nynorsk-nasjonalbiblioteknorm \
  --preserve-unreviewed \
  --sealed-cache-output /path/to/nynorsk_run_sealed_cache.yaml

# 3. Stream directly from a Hugging Face dataset repository:
normform batch \
  -i hf://NbAiLab/npsc_orto \
  --dataset-split train \
  -o npsc_standardized.jsonl \
  --lang nob \
  --style redaksjonsnorm \
  --limit 10000

Python API for Batch Standardization

from pathlib import Path
from normform import run_batch_standardization

run_batch_standardization(
    input_source="/path/to/corpus.jsonl",
    output_path=Path("/path/to/corpus_standardized.jsonl"),
    sealed_cache_output=Path("/path/to/sealed_cache.yaml"),
    lang="nob",
    style="redaksjonsnorm",
    preserve_unreviewed=False,
    text_field="text",
    output_field="normalized",
    batch_size=512,
    chunk_size=4096,
    device="cuda:0",
)

4.3 Generating [alt1|alt2] Morphological Alternatives (normform alternatives)

normform alternatives expands each sentence into alphabetically sorted [alt1|alt2] brackets while keeping all non-alternative text unchanged. This output is directly compatible with altmetrics and altwer for ASR/TTS evaluation where any valid spelling in the bracket is accepted as correct.

# Single sentence:
normform alternatives -t "Jenta kasta ballen." --lang nob
# -> [Jenta|Jenten] [kasta|kastet] ballen.

# Whole file (.txt or .jsonl):
normform alternatives -i references.jsonl -o references_with_alternatives.jsonl --lang nob

4.4 Auditing Inconsistencies & Change Statistics (normform audit)

normform audit supports three complementary auditing workflows:

A. Raw Source Text Inconsistency Audit (--raw)

Checks whether a raw document or corpus mixes competing morphological variants of the same candidate group (e.g., uses boka 14 times and boken 3 times in the same text):

normform audit -i manuscript.txt --lang nob --raw

B. Model & Style Consistency Audit (--uncached)

Tests whether the neural model (under a given --style or --prompt) makes conflicting choices for the same candidate group across different sentences when cross-sentence caching is disabled:

normform audit -i evaluation_set.jsonl --lang nob -p "boka,jenta,kasta" --uncached -o audit.json

C. Post-Standardization Corpus Change Statistics (--stats)

Compares the source --text-field and standardized --normalized-field in a processed .jsonl or .parquet corpus and outputs a ranked YAML report of the top morphological groups that caused changes:

normform audit -i corpus_standardized.jsonl --stats --top 50 -o change_report.yaml

Example output:

corpus: corpus_standardized.jsonl
total_records_scanned: 50000
total_records_changed: 14210 (28.42%)
unique_alternative_groups_changed: 842
top_alternatives:
  fram | frem: {chosen: frem, total_changes: 1204, changed_from: {fram: 1204}}
  boka | boken: {chosen: boken, total_changes: 618, changed_from: {boka: 618}}

4.5 Extracting Steering Style Words (normform extract-style)

Given an existing reference document (for example, an editorial style sample or an author's manuscript), normform extract-style discovers the top $N$ single-word morphological forms that best explain the document's empirical style choices via Information-Theoretic Log-Likelihood Submodular Maximization:

# Output full JSON report with Explained Style Variance (%):
normform extract-style -i reference_article.txt --lang nob -n 10

# Output only the comma-separated prompt string for direct piping:
PROMPTS=$(normform extract-style -i reference_article.txt --lang nob -n 10 --prompt-only)
normform standardize -i new_article.txt -p "$PROMPTS" --lang nob

4.6 Managing & Customizing Styles (normform styles)

NormForm decouples code from linguistic styles. Every style profile is a self-contained YAML file (style.yaml) hosted in its own Hugging Face dataset repository (1 repo = 1 style).

# 1. List built-in official styles and short aliases:
normform styles list

# 2. Download an official style to a local YAML file to inspect or customize:
normform styles pull redaksjonsnorm -o my_style.yaml

# 3. Edit my_style.yaml in any text editor, then push to your own HF dataset repo:
#    (Creates a private HF dataset repo by default; add --public to make it public)
normform styles push -f my_style.yaml --repo username/my-normform-style --lang nob

# 4. Use your custom style from anywhere (or pin a specific commit with @<sha>):
normform standardize -i document.txt --style username/my-normform-style
normform batch -i corpus.jsonl --style username/my-normform-style@main

5. Style File Format (style.yaml)

A NormForm style.yaml file has three human-readable sections:

_metadata:
  language: nob
  style_name: my-custom-style
  prompt_words:
    - boken
    - kastet
    - frem
    - syk

choices:
  # Mapping from "alt1 | alt2" (lowercase, alphabetical) to the chosen lowercase form:
  boka | boken: boken
  fram | frem: frem
  kasta | kastet: kastet
  "no | nå": nå

invalid_pairs:
  # Ambiguous or homographic candidate groups that must NEVER be altered:
  bad | badet: Homographic noun/verb inflection; preserve source token.
  verken | verket: Distinct lemmas (conjunction 'verken' vs. noun 'verket').

6. Official Models & Style Repositories

All official models and style repositories are publicly available on Hugging Face under NbAiLab and pinned by commit SHA for reproducibility:

7. License & Credits

Licensed under the Apache License 2.0 (see LICENSE).

Created by Per Egil Kummervold and Magnus Breder Birkenes at the National Library of Norway (Nasjonalbiblioteket).