PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-Lingual Semantic Role Labeling with High-Quality Translated Training Corpus

Hao Fei, Meishan Zhang, Donghong Ji

arXiv:2004.06295v2cs.CL

TL;DR

Low-resource languages may lack annotated SRL data, leaving competitive performance challenging. The paper translates gold-standard source annotations to build target-language training corpora and reports effective cross-lingual SRL results across seven languages and both transfer settings.

  • Problem

    Low-resource languages may lack annotated SRL datasets, making competitive SRL performance challenging.

    Source: §1 · p. 1

  • Method

    The paper translates gold-standard source SRL annotations into target-language corpora, combines them with source data, and uses multilingual representations with PGN-BiLSTM.

    Source: §1 · p. 2

  • Results

    Translation-based SRL is highly effective on UPB v1.0 across seven languages in single- and multi-source transfer, with further improvement from PGN-BiLSTM.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    Translated pseudo-corpora improve target-language SRL performance in the evaluated settings.

    Source: §1 · p. 2, §6 · p. 9

    The study focuses on unsupervised cross-lingual SRL.

    Source: §5.2 · p. 5

Abstract

from arXiv · show

Many efforts of research are devoted to semantic role labeling (SRL) which is crucial for natural language understanding. Supervised approaches have achieved impressing performances when large-scale corpora are available for resource-rich languages such as English. While for the low-resource languages with no annotated SRL dataset, it is still challenging to obtain competitive performances. Cross-lingual SRL is one promising way to address the problem, which has achieved great advances with the help of model transferring and annotation projection. In this paper, we propose a novel alternative based on corpus translation, constructing high-quality training datasets for the target languages from the source gold-standard SRL annotations. Experimental results on Universal Proposition Bank show that the translation-based method is highly effective, and the automatic pseudo datasets can improve the target-language SRL performances significantly.

1 Introduction

Cross-lingual SRL addresses the scarcity of labeled data in target languages. This paper translates gold-standard source training data into target-language training corpora and reports effective results across single- and multi-source settings, further improved with PGN-BiLSTM.

1 Introduction

  • SRL captures sentence-level meaning such as who did what to whom, while labeled data is abundant for English but scarce or unavailable for many target languages.The paper motivates cross-lingual SRL as a way to transfer resources from languages such as English.Source: §1 · p. 1
  • Prior cross-lingual SRL approaches use model transfer through language-independent features or project source annotations across parallel corpora and word alignments.Annotation projection can also be combined with model transfer.Source: §1 · p. 1
  • Translation-based SRL addresses noise from automatically annotated source data by translating gold-standard training data directly into the target language.The passage identifies neural machine translation advances as a reason this approach has potential for cross-lingual transfer.Source: §1 · pp. 1–2
  • The method translates gold-standard source sentences, projects their SRL annotations onto the target side, and trains a target SRL model on the resulting corpus.The source gold corpus is also merged with the translated target corpus; PGN enhances the BiLSTM to capture language differences.Source: §1 · p. 2
  • The translation-based method is highly effective in both single-source and multi-source SRL transfer, with further gains when PGN-BiLSTM is used.Experiments use the Universal Proposition Bank across seven languages.Source: §1 · p. 2
  • The paper presents translation-based corpus construction as its first approach to unsupervised cross-lingual SRL and evaluates the resulting pseudo-corpus across settings.The contribution statement characterizes the target-language pseudo training corpus as high-quality.Source: §1 · p. 2

2 Related Work

Prior cross-lingual transfer work mainly uses model transfer or annotation projection, while this work draws on translated treebanks and combines translation with model transfer for SRL.

2 Related Work

  • Cross-lingual transfer commonly uses model transfer or annotation projection, with projection creating target training instances from source annotations and parallel sentences.The paper situates its approach in translation-based treebank work and combines it with model transfer, a less explored combination for cross-lingual SRL.Source: §2 · p. 2
  • For cross-lingual SRL, prior projection work improves projection quality, jointly learns syntax and semantics, bootstraps to reduce noise, or jointly models translation and SRL.The cited examples include Tiedemann (2015), Kozhevnikov and Titov (2013), Akbik et al. (2015), and Aminian et al. (2019).Source: §2 · p. 2
  • The study uses UPB, whose consistent dependency-based universal SRL annotations follow English Proposition Bank frame and role schemes.UPB is built on the Universal Dependency Treebank and Proposition Bank.Source: §2 · p. 3
  • The work also uses multilingual contextualized representations with a PGN-BiLSTM model for cross-lingual SRL.The paper describes this model as capable of state-of-the-art cross-lingual SRL performance.Source: §2 · p. 3

3 SRL Translation

The method translates gold-standard source SRL data, then projects predicates and arguments into the target language. It uses alignment and POS information to score projections, resolve conflicts, and filter low-confidence annotations.

3 SRL Translation

  • The method translates gold-standard source SRL sentences and projects their predicates and arguments to form an automatic target-language corpus.Unlike annotation projection from automatically labeled source sentences, it begins with gold-standard source annotations.Source: §3 · p. 3
  • For each annotated source word, the method selects the target word with the highest alignment probability and scores the projection using alignment and target POS-tag consistency.Alignment probabilities come from a word-alignment tool; target POS-tag distributions come from a supervised target-language POS tagger.Source: §3 · pp. 3–4
  • When multiple source words align to one target word, predicate projections take precedence; otherwise, the highest-confidence projection is retained.The paper illustrates predicate–argument and argument–argument conflicts.Source: §3 · p. 4
  • A confidence threshold removes low-confidence projections; dropping a predicate projection also removes all its roles, while dropping an argument removes only that argument.The filtering threshold is denoted α.Source: §3 · p. 4

4 The SRL Model

The SRL model treats role labeling for a predicate as sequence labeling with a BiLSTM-CRF, adapted for multilingual inputs through language-aware parameters. It represents words with lexical, POS, and predicate cues, encodes them with PGN-BiLSTM, and predicts role-tag sequences with CRFs.

4 The SRL Model

  • The model frames dependency-based SRL for a given predicate as sequence labeling and adapts a multi-layer BiLSTM-CRF for multilingual inputs with a PGN module.The paper cites prior work for the task framing, model, and PGN module.Source: §4 · p. 4

4.1 Word Representation

  • Each word combines word-form, universal POS, and predicate-indicator features; the paper compares multilingual embeddings, ELMo, and BERT as word-form representations.BERT word representations average the inner-word piece vectors.Source: §4.1 · p. 4

4.2 Encoding Layer

  • PGN-BiLSTM uses a language embedding to dynamically generate language-aware BiLSTM parameters, including input, forget, output-gate, and cell-module weights.The encoder is a 3-layer bidirectional PGN-LSTM; the language embedding is an extra input used to obtain its parameters.Source: §4.2 · pp. 4–5

4.3 Output Layer

  • The model uses CRFs to score candidate SRL tag sequences from encoder outputs, with Viterbi decoding to select the highest-probability sequence.The model is a multi-layer BiLSTM-CRF for dependency-based SRL, recognizing semantic roles for a given predicate.Source: §4 · p. 4, §4.3 · p. 5

5 Experiments

The experiments evaluate translation-based cross-lingual SRL on seven UPB languages, then compare translated, source, and mixed training data across transfer settings. Translated data improves over source-only training, and mixing corpora with a language-aware PGN encoder yields further gains; multi-source transfer also outperforms single-source transfer.

5.1 Universal Proposition Bank

5.2 SRL Translation

  • The unsupervised setting assumes no gold-standard target-language SRL corpus and builds pseudo-training data by translating gold-standard source SRL data.Google Translate produces translations, fastAlign supplies word alignments, and supervised monolingual POS taggers provide universal POS tags.Source: §5.2 · p. 5

5.3 Settings

  • The experiments compare multilingual Emb, ELMo, and BERT representations and several SRL models, using α=0.4 as the universal projection threshold.BASIC is the vanilla BiLSTM baseline; MoE and MAN-MoE are stronger multilingual baselines, and PGN is evaluated as a language-aware encoder.Source: §5.3 · pp. 5–6, Figure 5 · p. 6

5.4 Cross-Lingual Transfer from English

  • Adding translated target data to source data yields a further average improvement of 4.7 over translated-only training, using BASIC for the comparison.The authors report similar gains when α=0, without sentence filtering.Source: §5.4 · p. 7
  • 60.3 versus 55.8 averaged F1: PGN improves SRC & TGT transfer by 4.5, while translated-only TGT gains 6.7 over SRC (51.1 versus 44.4).Source: Table 2 · p. 6, §5.4 · p. 7

5.5 Multi-Source Transfer

  • Multi-source transfer outperforms single-source transfer in all cases, and same-family language pairs often benefit one another.The reported family groupings include EN–DE and FR–IT–ES–PT; multilingual ELMo performs best, and the mixed corpus with PGN gives the strongest overall results.Source: §5.5 · pp. 7–8, Table 3 · p. 7, Table 4 · p. 7

5.6 Analysis

  • Across the analyzed roles, A0 has the highest F1, followed by A1, while A2 and AM-TMP perform slightly worse; multi-source transfer ranks best across roles and languages.The authors relate role-level differences partly to label frequencies and the transferability of notional words.Source: §5.6 · p. 8, Figure 7 · p. 8
  • SRL F1 decreases as the surface distance between predicates and arguments grows, while model rankings follow the overall results.Source: §5.6 · p. 9, Figure 8 · p. 9

6 Conclusion

The paper proposes corpus translation as an alternative for cross-lingual SRL, constructing target-language datasets from source gold-standard annotations. Experiments on UPB v1.0 report improvements across selected languages in both single-source and multi-source transfer.

6 Conclusion

  • The translation-based method constructs target-language SRL datasets by translating source gold-standard annotations.Source: §6 · p. 9
  • Combining gold-standard source corpora with pseudo translated target corpora is intended to enhance cross-lingual SRL models.Source: §6 · p. 9
  • On UPB v1.0, translated datasets significantly improved performance for all selected languages in both single-source and multi-source transfer.Source: §6 · p. 9
  • The study also investigates multilingual word representations and presents a PGN-BiLSTM encoder to better exploit mixed-language corpora.Source: §6 · p. 9

A Bilingual Transfer by Each Source

The section examines bilingual SRL transfer separately for each source-target language pair and directs readers to Table 5 for detailed results.

A Bilingual Transfer by Each Source

B Extended SRL Analysis

The extended analysis examines detailed role labeling and predicate-argument distance for Italian, Spanish, and Portuguese, with results presented in Figures 9 and 10.

B Extended SRL Analysis

Loading 2004.06295v2…