Cross-Lingual Transfer of Semantic Roles: From Raw Text to Semantic Roles

Maryam Aminian, Mohammad Sadegh Rasooli, Mona T. Diab

TL;DR

Most cross-lingual transfer methods for Semantic Role Labeling (SRL) assume that the target language already has reliable supervised tools like part-of-speech taggers, lemmatizers, and syntactic parsers. This paper shows that you can successfully project SRL annotations across parallel text to a target language relying solely on raw text—substituting supervised linguistic features with character-level BiLSTMs and unsupervised stem embeddings. The resulting lightweight neural model outperforms a feature-rich, supervised baseline across six out of seven target languages in the Universal Proposition Bank.

What Problem Does This Paper Address?

Semantic Role Labeling (SRL) determines "who did what to whom" by tagging predicates and their associated semantic arguments (e.g., Agent, Patient, Manner). While annotated SRL corpora exist for high-resource languages like English, building them for new languages is labor-intensive.

A common workaround is annotation projection: running an SRL model on the source side of a parallel corpus and projecting labels across word alignments to the target side. However, previous transfer frameworks rely on a pipeline of supervised target-language tools (accurate lemmatizers, POS taggers, and dependency parsers). For truly low-resource languages, these tools do not exist or perform poorly. The authors tackle a critical question: Can we build an accurate cross-lingual dependency-based SRL system when the target language has nothing more than raw parallel text?

Methods

The authors adapt the deep, syntax-agnostic SRL architecture of Marcheggiani et al. (2017) to function entirely without target-language supervised linguistic tools.

Source Sentence (English) ──[Supervised SRL]──> Source Semantic Roles
                                                     │
                                            (Word Alignments)
                                                     ▼
Target Sentence (Raw Text) ──[Projection Filter]─> Projected Roles
                                                     │
                                         [Train End-to-End]
                                                     ▼
                  Target SRL Model (Char-BiLSTMs + Word Embeddings)
  1. Annotation Projection and Density Filtering:
  2. English sentences are tagged using a supervised SRL model (trained on PropBank 3 converted to dependency heads).
  3. Roles are projected across word alignments produced by Giza++.
  4. To reduce noise from unaligned tokens, sentence pairs with alignment density below a threshold (80% for European languages, 60% for Chinese) are pruned.
  5. Feature-Free Target Architecture:
  6. No Supervised POS: Word tokens are represented by concatenating learned embeddings, pre-trained Word2Vec embeddings, and a character-level BiLSTM representation that captures subword and morphological cues.
  7. Predicate Modeling without Lemmatizers: State-of-the-art decoders depend heavily on predicate lemma identity. The authors test two alternatives to supervised lemmatizers:
  8. Character BiLSTMs (CModel): Distinct character-level BiLSTMs encode the predicate surface form for both the input encoder and the role-prediction decoder.
  9. Unsupervised Stems (UStem): Stems extracted via an unsupervised morphological analyzer (Morfessor FlatCat + a finite-state machine).
  10. Predicate Disambiguation: Traditional pipelines train separate sense classifiers for each predicate lemma. In the absence of lemmatization, the authors train a single 3-layer BiLSTM classifier with a softmax layer across all predicates in the language.

Datasets

  • Parallel Corpora (Training): Europarl (for German, Spanish, Finnish, French, Italian, and Portuguese) and MultiUN (2 million sampled sentence pairs for Chinese).
  • Embeddings: Structured skip-gram embeddings for English; Word2Vec trained on Wikipedia for target languages.
  • Evaluation Benchmark: Universal Proposition Banks (Universal PropBank) test sets across seven languages (German, Spanish, Finnish, French, Italian, Portuguese, and Chinese). Universal PropBank maps predicate frames across languages to standardized English Proposition Bank 3.0 roles.

Key Insights and Ideas

  • Character representations make POS taggers redundant: Subword character BiLSTMs effectively replace explicit POS features in predicting argument roles without loss in accuracy.
  • Unsupervised morphology can beat supervised tools: On morphologically complex languages like Finnish, using unsupervised stems (UStem: 64.5 F1) or character encoders (CModel: 64.6 F1) outperformed supervised UDPipe lemmas (SLem: 62.5 F1), showing that imperfect supervised tools can inject brittle errors into low-resource pipelines.
  • Scalability through architectural simplicity: Heavy feature-based bootstrapping baselines (e.g., Aminian et al., 2017) struggle to scale past tens of thousands of sentences due to memory constraints. The neural character-based architecture scales cleanly to training sets containing hundreds of thousands of projected sentences.

Key Results

The character-only model (CModel) was evaluated against the feature-heavy bootstrapping baseline of Aminian et al. (2017) (Bootstrap), which uses supervised lemmas, POS tags, and dependency parse trees.

Metric / Model German (de) Spanish (es) Finnish (fi) French (fr) Italian (it) Portuguese (pt) Chinese (zh)
Bootstrap (Gold Sense) 59.8 60.6 59.0 71.0 59.2 61.2 50.3
CModel (Gold Sense) 61.0 62.5 64.6 65.1 61.0 62.9 52.7
Bootstrap (Auto Sense) 55.0 52.2 53.1 63.4 52.3 53.9 42.5
CModel (Auto Sense) 57.0 56.0 58.9 58.5 55.5 56.5 42.7
  • Outperforms the supervised-feature baseline: Despite having zero access to syntactic trees or POS tags in the target language, CModel beats the Bootstrap baseline on 6 out of 7 languages.
  • Comparable to supervised lemmatization: CModel achieves performance parity with models trained using supervised UDPipe lemmatizers (SLem).

Limitations and Future Work

  • The Sense Disambiguation Bottleneck: Automatic predicate disambiguation causes a steep performance drop across all models (4 to 10 F1 points). Without lemmatization, the model cannot rely on lemma-specific sense priors (e.g., defaulting to .01), leading to poor generalization on out-of-vocabulary predicates.
  • French Underperformance: French was the only language where the system fell behind the baseline (65.1 vs. 71.0 F1). Error analysis revealed that the model failed to predict core arguments, tagging $A_0$ and $A_1$ as NULL in 20% and 30% of cases, respectively.
  • Parallel Data Dependency: The technique still requires large, aligned parallel corpora. The authors suggest investigating hybrid approaches that combine annotation projection with direct model transfer to support scenarios where even parallel text is scarce.

Summary updated

Abstract

We describe a transfer method based on annotation projection to develop a dependency-based semantic role labeling system for languages for which no supervised linguistic information other than parallel data is available. Unlike previous work that presumes the availability of supervised features such as lemmas, part-of-speech tags, and dependency parse trees, we only make use of word and character features. Our deep model considers using character-based representations as well as unsupervised stem embeddings to alleviate the need for supervised features. Our experiments outperform a state-of-the-art method that uses supervised lexico-syntactic features on 6 out of 7 languages in the Universal Proposition Bank.
BibTeX
@inproceedings{aminian2019cross,
  author = {Maryam Aminian and Mohammad Sadegh Rasooli and Mona T. Diab},
  title = {{Cross-Lingual Transfer of Semantic Roles: From Raw Text to Semantic Roles}},
  booktitle = {IWCS},
  year = {2019},
  url = {https://openreview.net/forum?id=qmpWroJN5y},
}

Loading related content...

Figure