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

Maryam Aminian, Mohammad Sadegh Rasooli, Mona T. Diab

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.

Introduction

Semantic Role Labeling (SRL) is a critical task in natural language processing that identifies the predicate-argument structure of a sentence—answering "who did what to whom." While high-quality SRL systems exist for resource-rich languages like English, the vast majority of the world's languages lack the large-scale annotated datasets (like PropBank) required for supervised training. This work presents a method to bridge this gap using cross-lingual transfer from raw parallel text, eliminating the need for supervised linguistic features in the target language.

Annotation Projection and Methodology

The core of the approach is annotation projection. By taking a parallel corpus (e.g., English-German), an English SRL model labels the source side. These labels are then projected onto the target language using automatic word alignments.

Example of annotation projection from English to German, showing how semantic roles (A0, A1, etc.) are transferred across languages.

Unlike previous methods that rely on supervised tools like part-of-speech (POS) taggers or dependency parsers in the target language, this model is designed to be feature-agnostic. It utilizes:

  • Character-level BiLSTMs: To capture morphological information implicitly, replacing the need for explicit POS tags.
  • Unsupervised Stems: Using tools like Morfessor to identify root forms without a supervised lemmatizer.
  • Word Embeddings: Standard pre-trained vectors to represent lexical meaning.

This architecture allows the model to learn semantic roles directly from the projected labels and raw text, making it applicable even to languages where no prior NLP infrastructure exists.

Experimental Results

The researchers evaluated their model on the Universal Proposition Bank across seven languages: German (de), Spanish (es), Finnish (fi), French (fr), Italian (it), Portuguese (pt), and Chinese (zh). The proposed character-based model (CModel) and the version using unsupervised stems (UStem) were compared against a "Bootstrap" baseline that utilizes supervised lemmas, POS tags, and dependency trees.

Main results (F-scores) comparing the proposed models against baselines across seven languages.

As shown in the results table, the feature-agnostic models outperformed the feature-heavy baseline in 6 out of 7 languages. Notably, in morphologically rich languages like Finnish (fi), the unsupervised approach (UStem/CModel) showed significant gains over the supervised baseline. This suggests that relying on potentially error-prone supervised tools in low-resource settings can be a bottleneck, and that raw text features are sufficient for high-quality transfer.

Conclusion

This research demonstrates that supervised linguistic features are not a prerequisite for cross-lingual SRL transfer. By leveraging character-level representations and unsupervised morphology, the proposed system achieves state-of-the-art performance using only parallel corpora. This significantly lowers the barrier for developing semantic tools for the world's under-resourced languages.

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},
}
Figure