Cross-Lingual Parsing

Leverages multilingual language models like XLM-RoBERTa or mBERT to generate cross-lingually aligned word representations. These models are pre-trained on concatenated corpora from 100+ languages, creating a shared embedding space where syntactically similar words occupy similar vector positions regardless of language. This alignment allows a parser trained on English treebanks to recognize subject-verb-object structures in Swahili or Hindi without explicit translation.

Strips away language-specific lexical items and relies on universal morpho-syntactic features to generalize parsing rules. Instead of learning that 'the' is a determiner, the model learns that tokens with the Universal Dependencies POS tag DET followed by NOUN form a det dependency. Features include:

• Universal part-of-speech tags
• Morphological features (case, gender, number)
• Dependency relation labels from the UD framework This abstraction allows direct transfer of parsing logic across language families.

Employs a gradient reversal layer to train a parser that cannot distinguish between source and target language representations. A language discriminator network attempts to identify which language a token belongs to, while the encoder is trained to fool the discriminator. This adversarial objective forces the encoder to produce language-agnostic features, ensuring the downstream parsing head operates on representations that abstract away from surface language identity.

Uses parallel corpora and automatic word alignment tools to project syntactic annotations from a source language to a target language. Given an English sentence with gold dependency parses and its French translation, word alignments map English tokens to French tokens, transferring the dependency structure. The projected noisy annotations then serve as silver-standard training data for a target-language parser, often refined through self-training or co-training iterations.

Relies on the Universal Dependencies (UD) framework as a syntactic interlingua that defines 17 universal POS tags and 37 universal dependency relations. Because UD treebanks exist for 100+ languages using identical annotation standards, a parser can be trained on a concatenation of multiple source treebanks. The consistent nsubj, obj, and amod relations across languages mean the parser learns cross-linguistically valid syntactic rules rather than language-specific heuristics.

Implements a partially shared architecture where language-specific parameters capture idiosyncratic phenomena while shared parameters model universal syntax. The encoder's lower transformer layers may be shared across all languages to learn universal morphology, while higher layers or language-specific adapter modules handle language-particular word order variations. This design prevents negative transfer where high-resource language patterns overwrite crucial low-resource language distinctions.

The foundational framework that makes cross-lingual parsing possible. UD defines a cross-linguistically consistent set of dependency relations and part-of-speech tags.

• Provides standardized annotation across 100+ languages
• Enables parser transfer by aligning syntactic categories
• Uses the CoNLL-U format for uniform data representation
• Critical for training multilingual models like XLM-RoBERTa

Transformer-based encoders that produce language-agnostic contextual embeddings. These models are pre-trained on 100+ languages simultaneously, creating a shared representation space.

• XLM-RoBERTa is the current standard backbone for cross-lingual transfer
• Encodes syntactic structure implicitly through self-attention
• Enables zero-shot transfer: train a parser on English, apply to Hindi
• Subword tokenization handles morphological variation across scripts

The neural architecture introduced by Dozat and Manning that dominates modern dependency parsing. Its biaffine attention mechanism computes scores for all possible head-dependent pairs globally.

• Uses BiLSTM or Transformer encoders for word representations
• Biaffine layer predicts both arc existence and relation labels
• Achieves state-of-the-art Labeled Attachment Score (LAS) across languages
• Easily adapted to cross-lingual settings by swapping the encoder for multilingual BERT

The dominant paradigm for cross-lingual parsing due to its global optimization over all possible arcs. Unlike transition-based methods, graph-based parsers score the entire tree structure simultaneously.

• Uses the Chu-Liu/Edmonds algorithm for maximum spanning tree decoding
• Naturally handles non-projective dependencies common in morphologically rich languages
• Arc-factored models assume independence between edges for tractability
• Higher-order extensions incorporate sibling and grandparent features

The core technique where a parser trained exclusively on source language treebanks (typically English) is applied directly to a target language with no annotated data.

• Relies on multilingual encoders to align syntactic representations across languages
• Performance depends on typological similarity between source and target
• Works best when both languages share word order patterns (e.g., SVO vs SOV)
• Active research area: bridging the gap for low-resource and distant language pairs

The standard tab-separated format for representing annotated linguistic data in the Universal Dependencies framework. Every cross-lingual parser consumes and produces this format.

• 10 columns: ID, FORM, LEMMA, UPOS, XPOS, FEATS, HEAD, DEPREL, DEPS, MISC
• HEAD and DEPREL columns define the dependency structure
• Enables seamless data interchange between parser implementations
• Used for both training data and evaluation with the official UD evaluation script