span-aligner 0.2.0

A utility for aligning and mapping text spans between different text representations.

Span Projecting & Alignment

A utility for aligning and mapping text spans between different text representations, and projecting annotations across languages using semantic alignment.

Features

• Span Alignment: Sanitize boundaries, fuzzy match segments, map spans between text versions.
• Span Projection: Project annotations from a source text (e.g., English) to a target text (e.g., Dutch) using embeddings.

Installation

Install dependencies:

pip install span-aligner

Usage

The package span_aligner provides two main classes: SpanAligner and SpanProjector.

• SpanAligner: Uses regex and fuzzy search. It is highly efficient but restricted to monolingual tasks (same language). It serves as a strong baseline for correcting boundary offsets or mapping annotations between slightly different versions of a text.

• SpanProjector: Uses word embeddings (Transformers) to align tokens semantically. It supports cross-lingual projection and handles significant paraphrasing. However, it is computationally more expensive.

  • Complexity: The default mwmf (Max Weight Matching) algorithm has a complexity of O(n³), meaning execution time increases exponentially with text length.
  • Use Case: Use when languages differ or when textual differences are too great for fuzzy matching.

Optimization & Best Practices

To achieve the best results while managing computational cost, follow these guidelines:

1. Choose the Right Tool for the Job

If the source and target texts are in the same language, always start with SpanAligner. It is significantly faster and creates precise splits. Only switch to SpanProjector if fuzzy matching fails due to low textual overlap.

2. Manage Text Length (Chunking)

The SpanProjector (specifically with mwmf) struggles with very long sequences.

• Split Texts: Break documents into logical segments (e.g., paragraphs, decisions, list items) before projection.
• Project Locally: Align spans within their corresponding segments rather than projecting a small span against an entire document.

3. Select the Appropriate Algorithm

• mwmf (Max Weight Matching): The gold standard. Finds the globally optimal alignment but is slow. Use for final, high-quality output on segmented text.
• inter (Intersection): Much faster. Works excellently for short, distinct spans (e.g., named entities like persons, locations, dates) where context is less critical.
• itermax: A balanced heuristic that offers better speed than mwmf with comparable quality for many tasks.

4. Translation-Assisted Projection (Hybrid Approach)

If direct cross-lingual projection yields subpar results, consider an intermediate translation step to simplify the alignment task:

1. Translate Source: Use an LLM or NMT model to translate the annotated source text (or just the spans) into the target language.
2. Align Locally: Use SpanAligner (or SpanProjector with inter) to map the translated spans onto the actual target text.

Tip: The translation should mimic the vocabulary of the target text as closely as possible.

• Workflow: annotated_source + target_text → LLM → rough_translated_source → SpanAligner → final_annotated_target

Span Aligner

Utilities for exact and fuzzy span mapping.

Get Annotations from Tagged Text

Extract structured spans and entities from a string with inline tags.

from span_aligner import SpanAligner

tagged_input = "<administrative_body>Environmental Committee</administrative_body> discussed the <impact_location>central park</impact_location> renovation on <publication_date>2025-12-15</publication_date>."

ner_map = {
    "administrative_body": "ADMINISTRATIVE BODY",
    "publication_date": "PUBLICATION DATE",
    "impact_location": "PRIMARY LOCATION"
}

span_map ={
    "motivation" : "MOTIVATION"
}

annotations = SpanAligner.get_annotations_from_tagged_text(
    tagged_input,
    ner_map=ner_map,
    span_map=span_map
)

print(annotations["entities"])
# Output:
#[
#    {'start': 0, 'end': 23, 'text': 'Environmental Committee', 'labels': ['ADMINISTRATIVE BODY']},
#    {'start': 38, 'end': 50, 'text': 'central park', 'labels': ['PRIMARY LOCATION']},
#    {'start': 65, 'end': 75, 'text': '2025-12-15', 'labels': ['PUBLICATION DATE']}
#]

Rebuild Tagged Text

Reconstruct a string with XML-like tags from raw text and span/entity lists.

from span_aligner import SpanAligner

text = "On 2026-01-12, the Budget Committee finalized the annual report."
# Entities corresponding to 'ADMINISTRATIVE BODY' label (indices skip "the ")
entities = [{"start": 19, "end": 35, "labels": ["administrative_body"]}]

tagged, stats = SpanAligner.rebuild_tagged_text(text, entities=entities)
print(tagged)
# Output: On 2026-01-12, the <administrative_body>Budget Committee</administrative_body> finalized the annual report.

Map Tags to Original

Align annotated spans from a tagged string back to their positions in the original text, allowing for noisy text or translation differences.

from span_aligner import SpanAligner

original_text = "Budget Committee met on 2026-01-12 to view\n\n the central park prject."
tagged_text = "<administrative_body>Budget Committee</administrative_body> met on <publication_date>2026-01-12</publication_date> to review the <impact_location>central park</impact_location> project."

mapped_tagged_text = SpanAligner.map_tags_to_original(
    original_text=original_text,
    tagged_text=tagged_text,
    min_ratio=0.7
)
print(mapped_tagged_text)
# Output preserves original text errors:
# "<administrative_body>Budget Committee</administrative_body> met on <publication_date>2026-01-12</publication_date> to view
#  the <impact_location>central park</impact_location> prject."

Span Projector

Project annotations from one text to another using semantic alignment (e.g., cross-lingual projection).

The process begins by generating embeddings for both source and target texts, creating a similarity matrix, and finding the optimal set of alignment pairs. Several algorithms are implemented for this matching phase, including mwmf, inter, itermax, fwd, rev, greedy, and threshold.

Project En -> En (Identity/Paraphrase)

Project annotations to a similar text in the same language. Functions similar to the spanAligner with improved fuzzy matching.

from span_aligner import SpanProjector

# Initialize projector (uses BERT embeddings by default)
projector = SpanProjector(src_lang="en", tgt_lang="en")

src_text = "The <ent>cat</ent> \n\n sat. on the mat."
tgt_text = "The cat sat on the mat."

tagged_tgt, spans = projector.project_tagged_text(src_text, tgt_text)
print(tagged_tgt)
# Output: The <ent>cat</ent> sat on the mat.

Project En -> Nl (Cross-Lingual)

Project annotations from an English source text to a Dutch target translation.

from span_aligner import SpanProjector

# Initialize projector
projector = SpanProjector(src_lang="en", tgt_lang="nl")

src_text = """DECISION LIST <contextual_location>Municipality of Zele</contextual_location>
 <administrative_body>Standing Committee</administrative_body> | <contextual_date>June 28, 2021</contextual_date>
  <title>1. Acceptance of candidacies for the examination procedure coordinator of Welfare</title>
  <decision>Acceptance of candidacies for the examination procedure coordinator of Welfare</decision>
  <title>2. Establishment of valuation rules for the integrated entity Municipality and Public Social Welfare Center (OCMW)</title>
  <decision>Establishment of valuation rules for the integrated entity Municipality and OCMW</decision>"""

tgt_text = """BESLUITENLIJST Gemeente Zele Vast bureau | 28 juni 20211.
 1. Aanvaarden kandidaturen examenprocedure coördinator Welzijn
 Aanvaarden kandidaturen examenprocedure coördinator Welzijn
 2. Vaststelling waarderingsregels geïntegreerde entiteit Gemeente en OCMW
 Vaststelling waarderingsregels geïntegreerde entiteit Gemeente en OCMW"""

tagged_tgt, spans = projector.project_tagged_text(src_text, tgt_text)
print(tagged_tgt)
# Output: BESLUITENLIJST <contextual_location>Gemeente Zele</contextual_location>
# <administrative_body>Vast bureau</administrative_body> <contextual_date>| 28 juni 20211</contextual_date>.
# <title>1. Aanvaarden kandidaturen examenprocedure coördinator Welzijn
# Aanvaarden kandidaturen examenprocedure coördinator</title> Welzijn
# <title>2. Vaststelling waarderingsregels geïntegreerde entiteit Gemeente en OCMW</title>
# <decision>Vaststelling waarderingsregels geïntegreerde entiteit Gemeente en OCMW</decision>

Sentence Aligner

Low-level class for aligning tokens between two texts (sentences or paragraphs) using transformer embeddings. Based on the work of simalign but optimized for span mapping (partial alignment instead of full text) and customized for different embedding providers (Ollama, SaaS providers, Transformers, Sentence-Transformers).

Initialize Aligner

from span_aligner import SentenceAligner

# Use bert embeddings (default) with BPE tokenization
aligner = SentenceAligner(model="bert", token_type="bpe") 

text_src = "This is a simple test sentence for alignment."
text_tgt = "Dit is een eenvoudige testzin voor uitlijning."

Get Text Embeddings

Retrieve tokens and embedding vectors for a string.

tokens_src, vecs_src = aligner.get_text_embeddings(text_src)
print(f"Src tokens: {len(tokens_src)}, Vectors: {vecs_src.shape}")
# Output: Src tokens: 9, Vectors: (10, 768)

Align Partial Substring

Find the alignment of a specific substring from source to target.

# Align "simple test"
res_sub = aligner.align_texts_partial_substring(text_src, text_tgt, "simple test")
print(f"Src tokens in result: {[t.text for t in res_sub.src_tokens]}")
# Output: Src tokens in result: ['simple', 'test']

Configuration & Advanced Usage

Embedding Models

The model parameter supports common transformer models:

• "bert": bert-base-multilingual-cased (Default, robust multilingual performance)
• "xlmr": xlm-roberta-base (Strong cross-lingual transfer)
• "xlmr-large": xlm-roberta-large (Higher accuracy, more resource intensive)

# Use xlm-roberta-base
projector = SpanProjector(model="xlmr")

Matching Algorithms

The matching_method parameter controls how the token similarity matrix is converted into an alignment.

• "mwmf" (Max Weight Matching): Finds the global optimal independent edge set. Best quality, O(n³) complexity.
• "inter" (Intersection): Intersection of forward and backward attention. High precision, lower recall, very fast.
• "itermax" (Iterative Max): Heuristic iterative maximization. Good speed/quality balance.
• "greedy" (Greedy): Selects best matches greedily. Fast but local optimum.

# Trade accuracy for speed with 'inter'
projector = SpanProjector(matching_method="inter")

Tokenization: BPE vs Word

• token_type="bpe" (Recommended): Uses the transformer's subword tokenizer (e.g. WordPiece). Handles rare words better and aligns closer to the model's internal representation.
• token_type="word": Splits by whitespace/punctuation. Simpler, but can result in [UNK] tokens for transformers.