Labeled Attachment Score (LAS)

LAS is calculated as the ratio of tokens with both correct head and label to the total number of tokens, excluding the root.

• Formula: LAS = (Correct Head & Label Tokens) / (Total Tokens)
• Strictness: Requires exact match on both head index and relation type
• Range: 0% to 100%, with state-of-the-art parsers achieving >95% on English newswire
• Contrast: More stringent than Unlabeled Attachment Score (UAS), which only checks head correctness

The distinction between LAS and UAS isolates the parser's ability to assign correct grammatical relations.

• UAS: Measures only structural correctness—is the head token correct?
• LAS: Adds the semantic layer—is the dependency type (e.g., nsubj, dobj, amod) correct?
• Label Accuracy: Can be derived as LAS / UAS, revealing how often correct heads receive correct labels
• Diagnostic Value: A high UAS but low LAS indicates strong structure learning but poor relation classification

Standard evaluation follows the CoNLL shared task conventions to ensure fair comparison across parsers.

• Punctuation Exclusion: Punctuation tokens are typically excluded from scoring to prevent inflating results
• Root Handling: The artificial ROOT token (index 0) is not scored
• Multiword Tokens: In UD treebanks, only syntactic words are scored, not orthographic tokens
• Gold Segmentation: Evaluation assumes gold-standard tokenization; errors in token boundaries are not penalized

Modern neural parsers achieve high LAS scores, with performance varying significantly by language and genre.

• English Penn Treebank: Deep biaffine parsers with BERT embeddings reach ~96% LAS
• Universal Dependencies: Multilingual parsers like Stanza achieve 85-95% LAS depending on language
• Morphologically Rich Languages: Finnish, Turkish, and Korean typically score lower due to complex morphology
• Out-of-Domain Drop: Parsers trained on newswire can drop 5-10% LAS on social media or clinical text

LAS directly impacts downstream tasks that depend on accurate syntactic analysis.

• Relation Extraction: Incorrect dependency labels cause false semantic triples in knowledge base construction
• Coreference Resolution: Syntactic paths between mentions rely on correct head identification
• Sentiment Analysis: Negation scope detection depends on accurate neg dependency arcs
• Machine Translation: Syntactic reordering models degrade with attachment errors
• Error Propagation: A 2% LAS improvement can yield measurable gains in end-task accuracy

Analyzing LAS errors reveals systematic weaknesses in parser design and training data.

• Prepositional Phrase Attachment: The classic ambiguity—does "with a telescope" modify the verb or noun?
• Coordination Scope: Determining which conjuncts share a common head in "apples and oranges from Florida"
• Long-Distance Dependencies: Relative clauses and wh-movement create non-local arcs that are harder to predict
• Label Confusion: Frequent confusion between nsubj and nsubj:pass, or advmod and obl

The complementary metric to LAS that measures only the syntactic head assignment accuracy, ignoring relation labels. A token is correct if its head is correctly identified, regardless of the dependency type.

• Formula: (Correct Heads) / (Total Tokens)
• Relationship: LAS ≤ UAS always, since LAS adds the label constraint
• Diagnostic Value: A high UAS but low LAS indicates the parser finds the right structure but struggles with fine-grained relation classification

A cross-linguistically consistent framework that standardizes the dependency relation labels used in LAS calculation. UD defines 37 universal syntactic relations organized into a hierarchy.

• Core Arguments: nsubj, obj, iobj for subject and object relations
• Nominal Modifiers: nmod, amod, nummod for noun phrase structure
• Clausal Modifiers: acl, advcl for subordinate clauses
• Function Words: case, det, aux for grammatical markers

LAS evaluation typically uses the UD v2 relation inventory as the gold standard label set.

The standard tab-separated format for annotated data used in LAS evaluation. Each token occupies one line with 10 fields, where field 7 (HEAD) and field 8 (DEPREL) directly define the gold standard for LAS calculation.

• HEAD column: The zero-indexed ID of the syntactic governor (0 for root)
• DEPREL column: The Universal Dependencies relation label
• Evaluation Scripts: The official conll18_ud_eval.py script computes LAS by comparing predicted and gold HEAD+DEPREL pairs
• Token Alignment: Multi-word tokens and empty nodes require special handling during scoring

The neural architecture by Dozat & Manning (2017) that achieves state-of-the-art LAS scores across many languages. It uses deep biaffine attention to jointly predict heads and labels.

• Architecture: BiLSTM encoder → four MLPs → biaffine classifiers
• Arc Scoring: A biaffine transformation over head and dependent representations computes pairwise attachment scores
• Label Prediction: A separate biaffine classifier predicts the relation type for each candidate arc
• Typical LAS: 90%+ on English Penn Treebank, 85-95% on UD treebanks depending on language complexity

A deterministic parsing paradigm that builds dependency trees incrementally through a sequence of shift-reduce actions. LAS evaluation applies identically to both transition-based and graph-based parsers.

• Arc-Eager Strategy: Attaches dependents as soon as they are complete, enabling early relation labeling
• Error Propagation: Greedy action selection can cascade errors, reducing LAS on long sentences
• Dynamic Oracles: Training technique that improves LAS by exposing the parser to non-gold states during learning
• Speed Advantage: Linear time complexity makes transition-based parsers suitable for production where LAS trade-offs are acceptable

The handling of crossing dependency arcs that occur in languages with free word order. Standard LAS evaluation does not penalize non-projectivity, but achieving high LAS on non-projective structures requires specialized algorithms.

• Chu-Liu/Edmonds Algorithm: Finds the maximum spanning tree in directed graphs, enabling non-projective decoding
• Pseudo-Projective Parsing: Converts non-projective trees to projective ones with augmented labels, then restores crossings post-hoc
• Language Impact: Czech, Dutch, and Latin show significant non-projectivity; LAS on these languages benefits from non-projective algorithms
• Evaluation Note: LAS is computed on the original non-projective gold standard regardless of parsing strategy