span-aligner 0.1.0__py3-none-any.whl

span_aligner/__init__.py

@@ -0,0 +1,3 @@
+from .aligner import SpanAligner
+
+__all__ = ["SpanAligner"]

span_aligner/aligner.py

@@ -0,0 +1,1126 @@
+"""
+SpanAligner Module
+==================
+
+A utility module for aligning and mapping text spans between different text representations,
+particularly useful for Label Studio annotation compatibility.
+
+This module provides functionality to:
+- Sanitize span boundaries to avoid special characters
+- Find exact and fuzzy matches of text segments in original documents
+- Map spans from one text representation to another
+- Rebuild tagged text with nested annotations
+- Merge result objects containing span annotations
+
+Typical use case: When text has been modified (e.g., cleaned, translated) and annotations
+need to be realigned to the original or modified text.
+"""
+
+from __future__ import annotations
+
+import re
+from difflib import SequenceMatcher
+from typing import Dict, Any, List, Tuple, Optional, Union
+from rapidfuzz import fuzz
+
+# Span sanitization helper for Label Studio compatibility
+SPECIAL_CHARS = {"\n", "\r", "\t", " "}
+
+
+class SpanAligner:
+    """
+    A utility class for aligning text spans between different text representations.
+    
+    This class provides static methods for:
+    - Sanitizing span boundaries
+    - Finding exact and fuzzy text matches
+    - Mapping spans from extracted/modified text back to original text
+    - Rebuilding tagged text with proper nesting
+    - Merging annotation result objects
+    
+    All methods are static and the class serves as a namespace for related functionality.
+    
+    Example Usage:
+        >>> original = "Hello, World!"
+        >>> result_obj = {
+        ...     "spans": [{"start": 0, "end": 5, "text": "Hello", "labels": ["greeting"]}],
+        ...     "entities": [],
+        ...     "task": {"data": {"text": ""}}
+        ... }
+        >>> success, mapped = SpanAligner.map_spans_to_original(original, result_obj)
+    """
+    @staticmethod
+    def sanitize_span(text: str, start: int, end: int) -> tuple[int, int]:
+        """
+        Adjust start/end indices so they do not land on special characters.
+        
+        Moves start forward and end backward to avoid whitespace and control characters
+        at span boundaries, which is important for Label Studio compatibility.
+        
+        Args:
+            text: The text containing the span.
+            start: The starting index of the span (inclusive).
+            end: The ending index of the span (exclusive).
+            
+        Returns:
+            tuple[int, int]: A tuple of (sanitized_start, sanitized_end) indices.
+                Both values are clamped to [0, len(text)] and guaranteed to satisfy start <= end.
+                
+        Example:
+            >>> SpanAligner.sanitize_span("  Hello  ", 0, 9)
+            (2, 7)  # Removes leading/trailing spaces
+        """
+        n = len(text)
+        s = max(0, min(start, n))
+        e = max(0, min(end, n))
+
+        # Move start forward while on a special char and s < e
+        while s < e and s < n and text[s] in SPECIAL_CHARS:
+            s += 1
+        # Move end backward while on a special char and s < e
+        while s < e and e > 0 and text[e-1] in SPECIAL_CHARS:
+            e -= 1
+
+        return s, e
+
+    @staticmethod
+    def _sequence_similarity(a: str, b: str) -> float:
+        """
+        Calculate the similarity ratio between two strings using SequenceMatcher.
+        
+        Args:
+            a: First string to compare.
+            b: Second string to compare.
+            
+        Returns:
+            float: Similarity ratio between 0.0 and 1.0, where 1.0 means identical strings.
+                Returns 0.0 if either string is empty.
+        """
+        if not a or not b:
+            return 0.0
+        return SequenceMatcher(None, a, b).ratio()
+
+
+    @staticmethod
+    def _find_exact(original_text: str, segment: str) -> List[int]:
+        """
+        Find all exact occurrences of a segment within the original text.
+        
+        Args:
+            original_text: The text to search within.
+            segment: The exact substring to find.
+            
+        Returns:
+            List[int]: A list of starting indices where the segment was found.
+                Empty list if no matches found.
+                
+        Example:
+            >>> SpanAligner._find_exact("hello world hello", "hello")
+            [0, 12]
+        """
+        indices = []
+        start = 0
+        while True:
+            idx = original_text.find(segment, start)
+            if idx == -1:
+                break
+            indices.append(idx)
+            start = idx + 1
+        return indices
+
+    @staticmethod
+    def _best_fuzzy_in_window(
+            original_text: str,
+            segment: str,
+            start_hint: Optional[int],
+            max_search_slack: int = 20
+        ) -> Tuple[Optional[int], Optional[int], float]:
+            """
+            Find the best fuzzy match for a segment within a window around a hint position.
+            Uses RapidFuzz for performance and better fuzzy matching.
+            """
+            if not segment:
+                return None, None, 0.0
+
+            # Calculate window bounds
+            if start_hint is None:
+                left = 0
+                right = len(original_text)
+            else:
+                left = max(0, start_hint - max_search_slack)
+                # Add segment length + slack to the window end
+                right = min(len(original_text), start_hint + len(segment) + 2 * max_search_slack)
+                
+            window = original_text[left:right]
+
+            if not window:
+                return None, None, 0.0
+
+            best_ratio = 0.0
+            best_start = None
+            best_end = None
+            
+            seg_len = len(segment)
+            # Allow length variation (slack)
+            # We allow candidates to be +/- 20% length of segment, but at least +/- 5 chars
+            slack_len = max(5, int(seg_len * 0.2))
+            min_len = max(1, seg_len - slack_len)
+            max_len = seg_len + slack_len
+
+            len_window = len(window)
+
+            # Sliding window search
+            for i in range(len_window):
+                # Optimization: If the remaining window is shorter than min_len, stop
+                if i + min_len > len_window:
+                    break
+                
+                # Limit candidate end to avoid checking excessively long strings
+                end_limit = min(i + max_len, len_window) + 1
+                
+                for j in range(i + min_len, end_limit):
+                    candidate = window[i:j]
+                    
+                    # Use rapidfuzz ratio
+                    ratio = fuzz.ratio(segment, candidate) / 100.0
+                    
+                    if ratio > best_ratio:
+                        best_ratio = ratio
+                        best_start = left + i
+                        best_end = left + j
+                        
+                        if best_ratio == 1.0:
+                            break
+                if best_ratio == 1.0:
+                    break
+
+            if best_start is None:
+                return None, None, 0.0
+
+            # Sanitize using the original SpanAligner helper
+            best_start, best_end = SpanAligner.sanitize_span(original_text, best_start, best_end)
+        
+            return best_start, best_end, best_ratio
+
+    @staticmethod
+    def _regex_word_sequence(
+        original_text: str,
+        segment: str,
+        start_hint: Optional[int] = None,
+        max_search_slack: int = 20
+    ) -> Optional[Tuple[int, int]]:
+        """
+        Find a segment in the original text using regex-based word sequence matching.
+        
+        This method tokenizes the segment into words and punctuation, then builds a
+        tolerant regex pattern that allows for varying whitespace/separators between
+        tokens. This is useful for matching text that may have different formatting.
+        
+        Args:
+            original_text: The text to search within.
+            segment: The text segment to find (will be tokenized).
+            start_hint: Approximate starting position to prioritize searching around.
+                If provided, searches near this position first before falling back to
+                full text search.
+            max_search_slack: Maximum distance from start_hint to search.
+                Default is 20 characters.
+                
+        Returns:
+            Optional[Tuple[int, int]]: A tuple of (start, end) indices if found,
+                or (None, None) if no match found.
+                
+        Example:
+            >>> # Matches "hello world" even with different whitespace
+            >>> SpanAligner._regex_word_sequence("hello   world", "hello world")
+            (0, 13)
+        """
+        # Tokenize into words and punctuation, keeping punctuation tokens
+        # Words: one or more word chars; Punct: any single non-word, non-space char
+        tokens = re.findall(r"\w+|[^\w\s]", segment)
+        if not tokens:
+            return None, None
+
+        # Build a tolerant pattern that matches tokens in order, allowing non-word separators/newlines after each
+        # Keep punctuation characters explicitly in the pattern
+        escaped = list(map(re.escape, tokens))
+        pattern = r"(?s)" + r"\W*".join(escaped)
+        try:
+            regex = re.compile(pattern)
+        except re.error:
+            return None, None
+
+        # If we have a hint, first search in a bounded region around it
+        if isinstance(start_hint, int):
+            left = max(0, start_hint - max_search_slack)
+            # allow for extra room to the right in case of many separators
+            right = min(len(original_text), start_hint + max(max_search_slack, len(segment) * 2))
+            subset = original_text[left:right]
+            m = regex.search(subset)
+            if m:
+                return left + m.start(), left + m.end()
+
+        # Fallback: search entire text
+        m = regex.search(original_text)
+
+        if not m:
+            return None, None
+        return m.start(), m.end()
+
+    @staticmethod
+    def map_spans_to_original(
+        original_text: str,
+        result_obj: Dict[str, Any],
+        min_ratio: float = 0.90,
+        logging: bool = False,
+        max_dist: int = 20,
+        enable_fuzzy: bool = False,
+    ) -> Tuple[bool, Dict[str, Any]]:
+        """
+        Map spans from a result object to their positions in the original text.
+        
+        This is the main alignment method that attempts to find the correct positions
+        of annotated spans in the original text. It uses multiple strategies:
+        1. Exact matching (fastest, most reliable)
+        2. Regex-based word sequence matching (handles whitespace variations)
+        3. Fuzzy matching (optional, for handling minor text differences)
+        
+        Args:
+            original_text: The original/target text to map spans onto.
+            result_obj: A dictionary containing annotation data with the following structure:
+                {
+                    "spans": [{
+                        "start": int,      # Approximate start position
+                        "end": int,        # Approximate end position  
+                        "text": str,       # The text content of the span
+                        "labels": [str]    # List of label names
+                    }, ...],
+                    "entities": [...],     # Same structure as spans
+                    "task": {
+                        "data": {"text": str}  # Will be updated with original_text
+                    }
+                }
+            min_ratio: Minimum similarity ratio (0.0-1.0) for fuzzy matching.
+                Default is 0.90.
+            logging: If True, prints debug information during alignment.
+                Default is False.
+            max_dist: Maximum allowed distance between approximate and actual
+                start positions. Matches further than this are rejected.
+                Default is 20 characters.
+                
+        Returns:
+            Tuple[bool, Dict[str, Any]]: A tuple of:
+                - bool: True if all spans were successfully aligned, False otherwise.
+                - Dict: Updated result_obj with mapped spans. Each span now includes:
+                    - start/end: Mapped positions (or None if unmatched)
+                    - text: Matched text from original (or None if unmatched)
+                    - status: "exact", "regex", "fuzzy", or "unmatched"
+                    - similarity: Match similarity score (0.0-1.0)
+                    - detected: The cleaned segment text that was searched for
+                    - approx_start: Original approximate start position
+                    
+        Example:
+            >>> original = "Hello, World!"
+            >>> result = {
+            ...     "spans": [{"start": 0, "end": 5, "text": "Hello", "labels": ["greeting"]}],
+            ...     "entities": [],
+            ...     "task": {"data": {"text": ""}}
+            ... }
+            >>> success, mapped = SpanAligner.map_spans_to_original(original, result)
+            >>> success
+            True
+            >>> mapped["spans"][0]["status"]
+            'exact'
+        """
+        input_spans: List[Dict[str, Any]] = result_obj.get("spans", [])
+        input_entities: List[Dict[str, Any]] = result_obj.get("entities", [])
+
+
+        def realign(items: List[Dict[str, Any]], enable_fuzzy: bool = False) -> Tuple[bool,List[Dict[str, Any]]]:
+            mapped: List[Dict[str, Any]] = []
+            all_aligned = True
+            for span in items:
+                
+                approx_start = span.get("start", 0)
+                segment = span.get("text", "") or ""
+                labels = span.get("labels", [])
+                clean_segment = segment.strip()
+                chosen_end = None
+
+                if logging:
+                    print("\n\n\n=======NEW SPAN==============")
+                    print(f"label: {labels}")
+                    print(f"cleaned span: '{clean_segment}' from original segment: '{segment}'")
+                    
+
+                exact_indices = SpanAligner._find_exact(original_text, clean_segment)
+                chosen_start = None
+                similarity = 1.0 if exact_indices else 0.0
+                status = "unmatched"
+                
+                # First try exact match
+                if exact_indices:
+                    chosen_start = min(exact_indices, key=lambda i: abs(i - approx_start))
+                    chosen_end = chosen_start + len(clean_segment)
+                    status = "exact"
+                
+                # Next try regex-based word sequence match (search near approx_start for all words in order)
+                else:
+                    step = 20
+                    slacks = sorted(list(set(list(range(0, max_dist + 1, step)) + [max_dist])))
+                    
+                    # 1. Try Regex progressively
+                    for current_slack in slacks:
+                        regex_start, regex_end = SpanAligner._regex_word_sequence(original_text, clean_segment, start_hint=approx_start, max_search_slack=current_slack)
+
+                        if regex_start is not None:
+                            chosen_start = regex_start
+                            chosen_end = regex_end
+                            # similarity compared against the fully matched span
+                            similarity = SpanAligner._sequence_similarity(clean_segment, original_text[regex_start:regex_end])
+                            status = "regex"
+                            break
+
+                    # 2. Try fuzzy matching if regex failed
+                    if status == "unmatched" and enable_fuzzy:
+                        
+                        step = 20
+                        slacks = sorted(list(set(list(range(0, max_dist + 1, step)) + [max_dist])))
+
+                        for current_slack in slacks:
+                            fuzzy_start, fuzzy_end, fuzzy_ratio = SpanAligner._best_fuzzy_in_window(original_text, clean_segment, start_hint=approx_start, max_search_slack=current_slack)
+                            
+                            if fuzzy_start is not None and fuzzy_ratio >= min_ratio:
+                                chosen_start = fuzzy_start
+                                chosen_end = fuzzy_end
+                                similarity = fuzzy_ratio
+                                status = "fuzzy"
+                                break
+                
+                            
+
+                # Check distance threshold
+                if chosen_start is not None and abs(chosen_start - approx_start) > max_dist:
+                    if logging:
+                        print(f"Match rejected due to distance: {abs(chosen_start - approx_start)} > {max_dist}")
+                    chosen_start = None
+                    chosen_end = None
+                    status = "unmatched"
+
+                if chosen_start is not None:
+                    # Sanitize the mapped span to avoid leading/trailing special characters
+                    chosen_start, chosen_end = SpanAligner.sanitize_span(original_text, chosen_start, chosen_end)
+                    matched_text = original_text[chosen_start:chosen_end]
+                else:  
+                    if logging:   
+                        print("No match found")  
+
+                    matched_text = None
+                    all_aligned = False
+
+                if logging:
+                    print("=====================")
+                    print(f"span: {span} segment: {clean_segment}")
+                    print(f"status:{status} similarity: {similarity}")
+                    print(f"pre sanit: {(chosen_start,chosen_end, len(clean_segment))}" if chosen_start is not None else "pre sanit: None")
+                    if chosen_start is not None:
+                        print(f"updated positions: (start: {chosen_start}; end: {chosen_end})")
+                        print(f"Extracted in original: '{original_text[chosen_start:chosen_end]}'")
+                    
+                mapped.append({
+                    "start": chosen_start,
+                    "end": chosen_end,
+                    "text": matched_text,
+                    "labels": labels,
+                    "status": status,
+                    "similarity": round(similarity, 4),
+                    "detected": clean_segment,
+                    "approx_start": approx_start,
+                })
+            return all_aligned, mapped
+
+        updated = dict(result_obj)
+        all_spans_aligned, updated["spans"] = realign(input_spans, enable_fuzzy)
+        all_entities_aligned, updated["entities"] = realign(input_entities, enable_fuzzy)
+        updated["task"]["data"]["text"] = original_text
+        return all_spans_aligned and all_entities_aligned, updated
+
+
+    @staticmethod
+    def merge_result_objects(
+        base: Dict[str, Any],
+        addition: Dict[str, Any],
+        span_from_name: str,
+        ner_from_name: str
+    ) -> Dict[str, Any]:
+        """
+        Merge two result objects by combining their span and entity lists.
+        
+        Creates a new dictionary based on the base object, then appends spans and
+        entities from the addition object.
+        
+        Args:
+            base: The base result object to merge into.
+            addition: The result object to merge from.
+            span_from_name: The key name for span annotations (e.g., "spans", "segmentation").
+            ner_from_name: The key name for NER/entity annotations (e.g., "entities").
+            
+        Returns:
+            Dict[str, Any]: A new merged dictionary with combined spans and entities.
+                The base object is shallow-copied, so nested objects may still be shared.
+                
+        Example:
+            >>> base = {"spans": [{"text": "A"}], "entities": []}
+            >>> addition = {"spans": [{"text": "B"}], "entities": [{"text": "C"}]}
+            >>> merged = SpanAligner.merge_result_objects(base, addition, "spans", "entities")
+            >>> len(merged["spans"])
+            2
+        """
+        merged = dict(base)
+
+        addition_spans = addition.get(span_from_name, [])
+        base_spans = merged.get(span_from_name, [])
+        merged[span_from_name] = base_spans + addition_spans
+
+        addition_ner = addition.get(ner_from_name, [])
+        base_ner = merged.get(ner_from_name, [])
+        merged[ner_from_name] = base_ner + addition_ner
+        return merged
+
+    @staticmethod
+    def _invert_label_map(tag_to_label: Dict[str, str]) -> Dict[str, str]:
+        """
+        Invert a tag-to-label mapping to create a label-to-tag mapping.
+        
+        Args:
+            tag_to_label: A dictionary mapping tag names to label names.
+                Can be None, in which case an empty dict is returned.
+                
+        Returns:
+            Dict[str, str]: A dictionary mapping label names to tag names.
+            
+        Example:
+            >>> SpanAligner._invert_label_map({"loc": "Location", "per": "Person"})
+            {'Location': 'loc', 'Person': 'per'}
+        """
+        return {v: k for k, v in (tag_to_label or {}).items()}
+
+    @staticmethod
+    def _sanitize_label_to_tag(label: str) -> str:
+        """
+        Convert a human-readable label to a sanitized XML-safe tag name.
+        
+        Converts the label to lowercase, replaces spaces with underscores,
+        and removes any characters that are not alphanumeric or underscores.
+        
+        Args:
+            label: The human-readable label to convert.
+            
+        Returns:
+            str: A sanitized tag name suitable for use in XML/HTML tags.
+                Returns "span" if the result would be empty.
+                
+        Example:
+            >>> SpanAligner._sanitize_label_to_tag("My Label (Special)")
+            'my_label_special'
+        """
+        # Fallback: convert human label to tag-like form
+        tag = label.strip().lower().replace(" ", "_")
+        tag = re.sub(r"[^a-z0-9_]+", "_", tag).strip("_")
+        return tag or "span"
+
+
+    @staticmethod
+    def _format_annotations(task: Any) -> Dict[str, Any]:
+        """
+        Extract and format annotations from a Label Studio task object.
+        
+        Parses the first annotation from the task and categorizes the results
+        into classification choices, entities, and segmentation spans.
+        Falls back to predictions if no annotations are available.
+        
+        Args:
+            task: A Label Studio task object with an `annotations` attribute.
+                Expected structure:
+                    task.annotations = [{
+                        "result": [
+                            {"type": "choices", "from_name": "type", "value": {"choices": [...]}},
+                            {"type": "labels", "from_name": "entities", "value": {...}},
+                            {"type": "labels", "from_name": "segmentation", "value": {...}}
+                        ]
+                    }]
+                If annotations are empty, predictions with the same structure
+                will be used as a fallback.
+                    
+        Returns:
+            Dict[str, Any]: A dictionary with three keys:
+                - "classification": List of classification choices
+                - "entities": List of entity annotation values
+                - "segmentation": List of segmentation span values
+        """
+        # Try annotations first, fall back to predictions if empty
+        results = []
+        if task.annotations:
+            results = task.annotations[0].get("result", [])
+        
+        # If no annotations, try predictions
+        if not results and hasattr(task, 'predictions') and task.predictions:
+            results = task.predictions[0].result or []
+        
+        classification = []
+        entities = []
+        spans = []
+
+        for ann in results:
+            ann_type = ann.get("type")
+            from_name = ann.get("from_name")
+            value = ann.get("value", {})
+
+            if ann_type == "choices" and from_name == "type":
+                if choices := value.get("choices"):
+                    classification = choices
+            elif ann_type == "labels":
+                if from_name == "entities":
+                    entities.append(value)
+                elif from_name == "segmentation":
+                    spans.append(value)
+        
+        return {
+            "classification": classification,
+            "entities": entities,
+            "segmentation": spans
+        }
+
+
+    @staticmethod
+    def update_mapped_with_rebuilt(
+        original_text: str,
+        mapped: Dict[str, Any],
+        span_label_mapping: Optional[Dict[str, str]] = None,
+        ner_label_mapping: Optional[Dict[str, str]] = None,
+        overwrite: bool = True
+    ) -> Dict[str, Any]:
+        """
+        Update a mapped result object with rebuilt tagged text.
+        
+        Takes a mapped result object (output from map_spans_to_original) and
+        generates tagged text from its spans and entities, storing the result
+        in the task data.
+        
+        Args:
+            original_text: The original text to use for rebuilding tags.
+            mapped: A mapped result object containing:
+                - "spans": List of span annotations
+                - "entities": List of entity annotations
+                - "task": {"data": {...}} - Task data to update
+            span_label_mapping: Optional tag-to-label mapping for spans.
+                Will be inverted to create label-to-tag mapping.
+            ner_label_mapping: Optional tag-to-label mapping for NER entities.
+                Will be inverted to create label-to-tag mapping.
+            overwrite: If True, overwrites "tagged_text" in task data.
+                If False, stores result in "tagged_text_unified" instead.
+                When overwriting, the original tagged_text is preserved in
+                "tagged_text_original" if it exists.
+                
+        Returns:
+            Dict[str, Any]: The same mapped object (modified in place) with:
+                - task.data.tagged_text (or tagged_text_unified): The rebuilt tagged text
+                - task.data.tagged_text_original: Original tagged_text if overwritten
+                - task.data.rebuild_stats: Statistics from rebuild operation
+                
+        Example:
+            >>> mapped = {"spans": [...], "entities": [...], "task": {"data": {}}}
+            >>> updated = SpanAligner.update_mapped_with_rebuilt("Hello World", mapped)
+            >>> "tagged_text" in updated["task"]["data"]
+            True
+        """
+        data = mapped.get("task", {}).get("data", {})
+        # text = data.get("text", "")
+        label_to_tag = {}
+        label_to_tag.update(SpanAligner._invert_label_map(span_label_mapping or {}))
+        label_to_tag.update(SpanAligner._invert_label_map(ner_label_mapping or {}))
+
+        rebuilt, stats = SpanAligner.rebuild_tagged_text(
+            original_text,
+            mapped.get("spans", []),
+            mapped.get("entities", []),
+            label_to_tag=label_to_tag,
+        )
+
+        # Preserve original and write unified
+        if "tagged_text" in data and not data.get("tagged_text_original"):
+            data["tagged_text_original"] = data.get("tagged_text")
+        if overwrite:
+            data["tagged_text"] = rebuilt
+        else:
+            data["tagged_text_unified"] = rebuilt
+        data["rebuild_stats"] = stats
+        return mapped
+
+
+
+#### From tagged text to task
+    @staticmethod
+    def get_annotations_from_tagged_text(
+        result: Union[dict, str],
+        *,
+        include_attachments: bool = True,
+        span_map: Optional[Dict[str, str]] = None,
+        ner_map: Optional[Dict[str, str]] = None,
+        class_map: Optional[Dict[str, str]] = None,
+        allowed_tags: Optional[List[str]] = None,
+    ) -> dict:
+        """
+        Convert a tagged result (with inline XML-like tags) into structured annotations.
+
+        Extracts spans and entities from tagged text by removing tags and tracking
+        character offsets in the resulting plain text. Supports nested tags and
+        custom tag-to-label mappings.
+
+        Args:
+            result: Input dictionary with 'tagged_text' key, or the tagged text string itself.
+            include_attachments: Whether to include text content inside <attachment> tags.
+            span_map: Dictionary mapping tag names to span labels.
+            ner_map: Dictionary mapping tag names to entity labels.
+            class_map: Dictionary mapping document classifications to labels.
+            allowed_tags: List of tag names to process. If None, derived from map keys.
+
+        Returns:
+            dict: A dictionary containing:
+                - spans: List of span (segmentation) objects
+                - entities: List of entity (NER) objects
+                - plain_text: The text content with tags removed
+                - tagged_text: The original tagged text used
+                - document_classification: The classification from input result (if any)
+        
+        Notes:
+            - Spans are derived by removing tags while tracking character offsets in the
+              plain text. Nested tags are supported; spans may overlap.
+            - `span_map` lets you rename tags to match your LS label config.
+            - `allowed_tags` limits which tags are turned into spans. If None, uses the
+              tag set defined in your prompts.
+        """
+
+        # Resolve tagged_text input
+        tagged_text = ""
+        doc_class = None
+        if isinstance(result, dict):
+            tagged_text = result.get("tagged_text", "")
+            doc_class = result.get("document_classification")
+        else:
+            tagged_text = str(result or "")
+
+        if not tagged_text:
+            raise ValueError("No tagged_text found in input result.")
+
+        # Default allowed tags (from your SYSTEM/USER prompts)
+        if allowed_tags is None:
+            # Safely handle None maps
+            s_map = span_map or {}
+            n_map = ner_map or {}
+            allowed_tags = list(n_map.keys()) + list(s_map.keys())
+        
+        # Merge span_map and ner_map safely into annotation_map
+        annotation_map = {}
+        for mapping in (span_map, ner_map):
+            if mapping:
+                annotation_map.update(mapping)
+
+        # If annotation_map ends up empty, initialize with identity mapping
+        if not annotation_map:
+            annotation_map = {t: t for t in allowed_tags}
+
+        # Regex to capture bare tags like <tag> or </tag>
+        tag_re = re.compile(r"<(/?)([a-zA-Z_][a-zA-Z0-9_-]*)>")
+
+        plain_parts: List[str] = []
+        spans: List[dict] = []
+        entities: List[dict] = []
+
+        stack: List[Tuple[str, int]] = []  # (tag_name_lower, start_offset_in_plain)
+
+        pos_in = 0  # position in tagged_text
+        pos_out = 0  # position in plain text we are building
+
+        def emit_text(s: str):
+            nonlocal pos_out
+            if not s:
+                return
+            plain_parts.append(s)
+            pos_out += len(s)
+
+        # Attachment handling: if we skip attachments, when inside attachments or attachment, we don't emit text
+        inside_attachments_level = 0
+
+        for m in tag_re.finditer(tagged_text):
+            # Emit any literal text before this tag
+            literal = tagged_text[pos_in:m.start()]
+            current_tag_is_attachment = inside_attachments_level > 0
+
+            if include_attachments or not current_tag_is_attachment:
+                emit_text(literal)
+
+            is_closing = bool(m.group(1))
+            tag_name = m.group(2).lower()
+
+            # Track attachments nesting regardless of allowed_tags so we can drop their content when requested
+            if tag_name in ("attachments", "attachment"):
+                if not is_closing:
+                    inside_attachments_level += 1
+                else:
+                    inside_attachments_level = max(0, inside_attachments_level - 1)
+
+            # Handle span stack only for allowed tags
+            if tag_name in allowed_tags:
+                if not is_closing:
+                    # Opening tag
+                    stack.append((tag_name, pos_out))
+                else:
+                    # Closing tag — find the last matching opening tag
+                    # Iterate backwards to find the matching opening tag
+                    found_open = False
+                    for i in range(len(stack) - 1, -1, -1):
+                        open_tag, start_off = stack[i]
+                        if open_tag == tag_name:
+                            # Pop all tags above the matching one (handle mismatched nesting)
+                            stack = stack[:i]
+                            end_off = pos_out
+                            
+                            # Create a span only if it has positive length
+                            if end_off > start_off:
+                                full_span_text = ("".join(plain_parts))[start_off:end_off]
+                                
+                                # Adjust start to skip leading newlines
+                                adjusted_start = start_off
+                                span_text = full_span_text
+                                
+                                while span_text.startswith('\n'):
+                                    adjusted_start += 1
+                                    span_text = span_text[1:]
+                                
+                                # Adjust end to skip trailing newlines
+                                adjusted_end = end_off
+                                while span_text.endswith('\n'):
+                                    adjusted_end -= 1
+                                    span_text = span_text[:-1]
+                                
+                                # Only create span if there's content after trimming
+                                if adjusted_end > adjusted_start:
+                                    annotation_entry = {
+                                        "start": adjusted_start,
+                                        "end": adjusted_end,
+                                        "text": span_text,
+                                        "labels": [annotation_map.get(tag_name, tag_name)]
+                                    }
+                                    
+                                    if ner_map and tag_name in ner_map:
+                                        entities.append(annotation_entry)
+                                    else:
+                                        spans.append(annotation_entry)
+                            found_open = True
+                            break
+                    # If no matching opening tag found, ignore gracefully
+
+            pos_in = m.end()
+
+        # Emit remaining tail text
+        tail = tagged_text[pos_in:]
+        if include_attachments or inside_attachments_level == 0:
+            emit_text(tail if include_attachments else "")
+
+        plain_text = "".join(plain_parts)
+
+        return { 
+            "spans": spans, 
+            "entities": entities, 
+            "plain_text": plain_text,
+            "tagged_text": tagged_text,
+            "document_classification": doc_class
+        }
+
+    @staticmethod
+    def tagged_text_to_task(
+        result: Union[dict, str],
+        *,
+        include_attachments: bool = True,
+        span_map: Optional[Dict[str, str]] = None,
+        ner_map: Optional[Dict[str, str]] = None,
+        class_map: Optional[Dict[str, str]] = None,
+        allowed_tags: Optional[List[str]] = None,
+    ) -> dict:
+        """
+        Convert a tagged result into an uploader-ready Label Studio task.
+
+        Uses `get_annotations_from_tagged_text` to parse the input and formats
+        the output as expected by the Label Studio uploader class.
+
+        Args:
+            result: Input dictionary with 'tagged_text' key, or the tagged text string.
+            include_attachments: Whether to include text content inside <attachment> tags.
+            span_map: Dictionary mapping tag names to span labels.
+            ner_map: Dictionary mapping tag names to entity labels.
+            class_map: Dictionary mapping document classifications to labels.
+            allowed_tags: List of tag names to process. If None, derived from map keys.
+
+        Returns:
+            dict: A dictionary ready for Label Studio import, containing:
+                - task: Task data including text and metadata
+                - spans: Extracted spans
+                - entities: Extracted entities
+                - labels: Classification labels (if applicable)
+        """
+        # Parse annotations using shared logic
+        parsed = SpanAligner.get_annotations_from_tagged_text(
+            result,
+            include_attachments=include_attachments,
+            span_map=span_map,
+            ner_map=ner_map,
+            class_map=class_map,
+            allowed_tags=allowed_tags
+        )
+        
+        spans = parsed["spans"]
+        entities = parsed["entities"]
+        plain_text = parsed["plain_text"]
+        tagged_text = parsed["tagged_text"]
+        doc_class = parsed["document_classification"]
+
+        # Handle classification mapping
+        classification_labels = []
+        if doc_class and class_map and doc_class in class_map:
+            classification_labels = [class_map[doc_class]]
+
+        content = {
+            "task": {
+                "data": {
+                    "text": plain_text,
+                    "tagged_text": tagged_text,
+                    "meta": {
+                        "segments": len(spans),
+                        "labels_present": sorted({(s.get("labels") or [""])[0] for s in spans}),
+                        "include_attachments": include_attachments,
+                        "document_classification": doc_class or ""
+                    }
+                }
+            },
+            "spans": spans,
+            "labels": classification_labels,
+            "entities": entities
+        }
+
+        return content
+
+
+#### From task to tagged text
+    @staticmethod
+    def rebuild_tagged_text(
+        original_text: str,
+        spans: List[Dict[str, Any]],
+        entities: List[Dict[str, Any]],
+        label_to_tag: Optional[Dict[str, str]] = None
+    ) -> Tuple[str, Dict[str, int]]:
+        """
+        Rebuild text with nested XML-style tags from span and entity annotations.
+        
+        Creates properly nested tags from annotations, handling overlapping spans
+        by skipping crossing (non-nested) annotations to maintain valid XML structure.
+        
+        Args:
+            original_text: The source text to add tags to.
+            spans: List of span annotations, each with:
+                - "start": int - Starting character index
+                - "end": int - Ending character index (exclusive)
+                - "labels": List[str] - Label names (first one is used)
+            entities: List of entity annotations (same structure as spans).
+            label_to_tag: Optional mapping from label names to tag names.
+                If a label is not in the mapping, it will be sanitized to
+                create a valid tag name.
+                
+        Returns:
+            Tuple[str, Dict[str, int]]: A tuple of:
+                - str: The text with XML tags inserted (e.g., "<tag>text</tag>")
+                - Dict with statistics:
+                    - "total": Total number of valid annotations processed
+                    - "skipped_crossing": Number of annotations skipped due to
+                      crossing (non-nested) overlaps
+                      
+        Note:
+            - Annotations with invalid positions (negative, overlapping bounds,
+              or exceeding text length) are silently skipped.
+            - For overlapping annotations, outer (longer) spans are preferred.
+            - Crossing annotations that would create invalid XML are skipped.
+            
+        Example:
+            >>> text = "Hello World"
+            >>> spans = [{"start": 0, "end": 11, "labels": ["sentence"]}]
+            >>> entities = [{"start": 0, "end": 5, "labels": ["greeting"]}]
+            >>> result, stats = SpanAligner.rebuild_tagged_text(text, spans, entities)
+            >>> result
+            '<sentence><greeting>Hello</greeting> World</sentence>'
+        """
+        annotations: List[Dict[str, Any]] = []
+
+        def to_tag(lbls: List[str]) -> Optional[str]:
+            if not lbls:
+                return None
+            lbl = lbls[0]
+            if label_to_tag and lbl in label_to_tag:
+                return label_to_tag[lbl]
+            return SpanAligner._sanitize_label_to_tag(lbl)
+
+        def add_items(items: List[Dict[str, Any]]):
+            for it in items or []:
+                s = it.get("start")
+                e = it.get("end")
+                if not isinstance(s, int) or not isinstance(e, int) or s is None or e is None or s < 0 or e <= s or e > len(original_text):
+                    continue
+                tag = to_tag(it.get("labels") or [])
+                if not tag:
+                    continue
+                annotations.append({
+                    "start": s,
+                    "end": e,
+                    "tag": str(tag),
+                    "length": e - s,
+                })
+
+        add_items(spans)
+        add_items(entities)
+
+        # Sort: by start asc, longer first (end desc) to open outers before inners
+        annotations.sort(key=lambda a: (a["start"], -a["length"]))
+
+        # Index starts and ends
+        starts: Dict[int, List[Dict[str, Any]]] = {}
+        for a in annotations:
+            starts.setdefault(a["start"], []).append(a)
+        for pos in starts:
+            starts[pos].sort(key=lambda a: -a["length"])  # longer first
+
+        ends: Dict[int, List[Dict[str, Any]]] = {}
+        for a in annotations:
+            ends.setdefault(a["end"], []).append(a)
+
+        event_positions = sorted({0, len(original_text), *starts.keys(), *ends.keys()})
+
+        pieces: List[str] = []
+        stack: List[Dict[str, Any]] = []
+        last = 0
+        skipped_cross = 0
+
+        for pos in event_positions:
+            if pos > last:
+                pieces.append(original_text[last:pos])
+
+            # Close all tags that end here (LIFO)
+            while stack and stack[-1]["end"] == pos:
+                top = stack.pop()
+                pieces.append(f"</{top['tag']}>")
+
+            # Open tags that start here (outer first)
+            for ann in starts.get(pos, []):
+                # Crossing check: if an open tag exists with end < ann.end (not nested), skip ann
+                if stack and ann["end"] > stack[-1]["end"]:
+                    skipped_cross += 1
+                    continue
+                pieces.append(f"<{ann['tag']}>")
+                stack.append(ann)
+
+            last = pos
+
+        # Tail
+        pieces.append(original_text[last:])
+
+        # Close any still-open tags (best-effort)
+        while stack:
+            top = stack.pop()
+            pieces.append(f"</{top['tag']}>")
+
+        return "".join(pieces), {"total": len(annotations), "skipped_crossing": skipped_cross}
+
+    @staticmethod
+    def rebuild_tagged_text_from_task(task: Any, mapping: Dict[str, str]) -> str:
+        """
+        Generate tagged text from a Label Studio task's annotations.
+        
+        Extracts annotations from the task and rebuilds the text with XML-style
+        tags around annotated spans.
+        
+        Args:
+            task: A Label Studio task object with:
+                - task.annotations: List of annotation objects
+                - task.data: Dict containing "text" key with the source text
+            mapping: A dictionary mapping label names to tag names to use in
+                the output. Labels not in the mapping will be sanitized to
+                create tag names.
+                
+        Returns:
+            str: The text with XML-style tags inserted around annotated spans.
+            
+        Example:
+            >>> # Returns something like: "<greeting>Hello</greeting>, World!"
+        """
+        extracted = SpanAligner._format_annotations(task)
+        text = task.data.get("text", "")
+
+        retagged, _ = SpanAligner.rebuild_tagged_text(
+            text,
+            spans=extracted["segmentation"],
+            entities=extracted["entities"],
+            label_to_tag=mapping
+        )
+
+        return retagged
+    
+
+#### Transpose tags back to original text
+    @staticmethod
+    def map_tags_to_original(
+        original_text: str,
+        tagged_text: str, 
+        min_ratio: float = 0.8,
+        max_dist: int = 20,
+        logging: bool = False
+    ) -> str:
+        """
+        Map spans from tagged text back to their positions in the original text.
+        
+        Takes tagged text with XML-style tags and aligns the annotated spans
+        back to their positions in the provided original text. Uses exact,
+        regex-based, and fuzzy matching to find the best alignment.
+        
+        Args:
+            original_text: The original untagged text.
+            tagged_text: The text with XML-style tags indicating spans.
+            min_ratio: Minimum similarity ratio (0.0-1.0) for fuzzy matching.
+                Defaults to 0.8.
+            max_dist: Maximum character distance from approximate position
+                to consider a match valid. Defaults to 20.
+            logging: If True, prints detailed debug information during mapping.
+                Defaults to False.
+        """        # First, extract spans/entities from tagged_text
+        temp_content = SpanAligner.tagged_text_to_task(
+            tagged_text,
+            include_attachments=True,
+            allowed_tags=None  # allow all tags
+        )
+
+        result_obj = {
+            "spans": temp_content.get("spans", []),
+            "entities": temp_content.get("entities", []),
+            "task": {
+                "data": {
+                    "text": ""  # will be filled later
+                }
+            }
+        }
+
+        # Now map spans/entities back to original_text
+        success, mapped = SpanAligner.map_spans_to_original(
+            original_text,
+            result_obj,
+            min_ratio=min_ratio,
+            max_dist=max_dist,
+            logging=logging
+        )
+
+        return mapped["task"]["data"].get("tagged_text", "")

span_aligner-0.1.0.dist-info/METADATA

@@ -0,0 +1,122 @@
+Metadata-Version: 2.4
+Name: span-aligner
+Version: 0.1.0
+Summary: A utility for aligning and mapping text spans between different text representations.
+License: MIT
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: rapidfuzz>=3.0.0
+Provides-Extra: dev
+Requires-Dist: pytest>=7.0.0; extra == "dev"
+Dynamic: license-file
+
+# Span Aligner
+
+A utility for aligning and mapping text spans between different text representations, particularly useful for Label Studio annotation compatibility.
+
+## Features
+
+- Sanitize span boundaries to avoid special characters.
+- Find exact and fuzzy matches of text segments in original documents.
+- Map spans from one text representation to another.
+- Rebuild tagged text with nested annotations.
+- Merge result objects containing span annotations.
+
+## Installation
+
+Install from source:
+
+```bash
+pip install .
+```
+
+For development:
+
+```bash
+pip install -e ".[dev]"
+```
+
+## Usage
+
+```python
+from span_aligner import SpanAligner
+
+original = "Hello, World!"
+result_obj = {
+    "spans": [{"start": 0, "end": 5, "text": "Hello", "labels": ["greeting"]}],
+    "entities": [],
+    "task": {"data": {"text": ""}}
+}
+
+success, mapped = SpanAligner.map_spans_to_original(original, result_obj)
+print(mapped)
+```
+
+### Map Tags to Original
+
+Align annotated spans from a tagged string back to their positions in the original text, keeping the mistakes and original text as written in the original.
+
+```python
+original_text = "The quick brown fox jumps\n\n over the dog."
+# Imagine the text was slightly modified or translated, but tags are present
+tagged_text = "The <adj>quick</adj> brown fox jumps over the <animal>dog</animal>."
+
+mapped_tagged_text = SpanAligner.map_tags_to_original(
+    original_text=original_text,
+    tagged_text=tagged_text,
+    min_ratio=0.8
+)
+print(mapped_tagged_text)
+# Output might look like: "The <adj>quick</adj> brown fox jumps\n\n over the <animal>dog</animal>."
+# (If original text differed slightly, tags would be placed on best matching spans)
+```
+
+### Rebuild Tagged Text
+
+Reconstruct a string with XML-like tags from raw text and span/entity lists.
+
+```python
+text = "Hello World"
+spans = [{"start": 0, "end": 11, "labels": ["sentence"]}]
+entities = [{"start": 6, "end": 11, "labels": ["location"]}]
+
+tagged, stats = SpanAligner.rebuild_tagged_text(text, spans, entities)
+print(tagged)
+# Output: <sentence>Hello <location>World</location></sentence>
+```
+
+### Rebuild Tagged Text from Task
+
+Generate tagged text directly from a Label Studio task object.
+
+```python
+# Assuming 'task' is a Label Studio task object (or similar structure)
+# with .data['text'] and .annotations attributes
+mapping = {"Location": "loc", "Person": "per"}
+
+tagged_output = SpanAligner.rebuild_tagged_text_from_task(task, mapping)
+print(tagged_output)
+```
+
+### Get Annotations from Tagged Text
+
+Extract structured spans and entities from a string with inline tags.
+
+```python
+tagged_input = "Visit <loc>Paris</loc> and see the <landmark>Eiffel Tower</landmark>."
+
+annotations = SpanAligner.get_annotations_from_tagged_text(
+    tagged_input,
+    ner_map={"loc": "Location", "landmark": "Location"}
+)
+
+print(annotations["entities"])
+# Output: 
+# [
+#   {"start": 6, "end": 11, "text": "Paris", "labels": ["Location"]},
+#   {"start": 24, "end": 36, "text": "Eiffel Tower", "labels": ["Location"]}
+# ]
+print(annotations["plain_text"])
+# Output: "Visit Paris and see the Eiffel Tower."
+```

span_aligner-0.1.0.dist-info/RECORD

@@ -0,0 +1,7 @@
+span_aligner/__init__.py,sha256=ERLPBS6aad_17IPcoMDkVqi7lrSJxuHgqrdh69EN9xI,63
+span_aligner/aligner.py,sha256=HYo5CvmC9WrnsJo7bFFcUHoJaZhiNf4aZ7pOEzd2MlA,47658
+span_aligner-0.1.0.dist-info/licenses/LICENSE,sha256=TqCZNrAXPrgWq9k95te7bOOyXztjxjxWQWnsHSqT8SM,1096
+span_aligner-0.1.0.dist-info/METADATA,sha256=nqV2rIVPmFCR2aiZ3KanmFx_ZlYfShz6JMp1SVRHld8,3583
+span_aligner-0.1.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+span_aligner-0.1.0.dist-info/top_level.txt,sha256=syADug30Z0JSDJXnan6CIBWuI4mCdpghyDa48kj69VY,13
+span_aligner-0.1.0.dist-info/RECORD,,

span_aligner-0.1.0.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (80.9.0)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

span_aligner-0.1.0.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Stefaan Vercoutere
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

span_aligner-0.1.0.dist-info/top_level.txt

@@ -0,0 +1 @@
+span_aligner