[论文解读] Choosing Transfer Languages for Cross-Lingual Learning

Yu-Hsiang Lin, Chian-Yu Chen+11 more|arXiv (Cornell University)|May 29, 2019

Natural Language Processing Techniques参考文献 41被引用 33

一句话总结

论文把跨语言学习中的转移语言选择视为一个学习排序问题（LangRank），结合数据集统计与语言类型学，预测在给定低资源任务语言下，在 MT、EL、POS 标注和依存句法分析中的最佳高资源转移语言。

ABSTRACT

Cross-lingual transfer, where a high-resource transfer language is used to improve the accuracy of a low-resource task language, is now an invaluable tool for improving performance of natural language processing (NLP) on low-resource languages. However, given a particular task language, it is not clear which language to transfer from, and the standard strategy is to select languages based on ad hoc criteria, usually the intuition of the experimenter. Since a large number of features contribute to the success of cross-lingual transfer (including phylogenetic similarity, typological properties, lexical overlap, or size of available data), even the most enlightened experimenter rarely considers all these factors for the particular task at hand. In this paper, we consider this task of automatically selecting optimal transfer languages as a ranking problem, and build models that consider the aforementioned features to perform this prediction. In experiments on representative NLP tasks, we demonstrate that our model predicts good transfer languages much better than ad hoc baselines considering single features in isolation, and glean insights on what features are most informative for each different NLP tasks, which may inform future ad hoc selection even without use of our method. Code, data, and pre-trained models are available at https://github.com/neulab/langrank

研究动机与目标

• Motivate and formalize the problem of selecting optimal transfer languages for a given low-resource task language.
• Propose LangRank, a ranking model that uses dataset-dependent and dataset-independent features to predict transfer-language usefulness.
• Demonstrate that LangRank outperforms single-feature baselines across MT, EL, POS tagging, and dependency parsing.
• Analyze feature importance to provide insights for educated guesses even without full training data.

提出的方法

• Formulate transfer-language selection as a learning-to-rank problem over a set of candidate transfer languages for a given task language t.
• Extract features for each language pair (t,a) including dataset-dependent features (e.g., dataset size, Type-Token Ratio, word/subword overlap) and dataset-independent linguistic distances (e.g., genetic, syntactic, phonological, geographic, inventory, featural distances from URIEL).
• Train a gradient boosted decision trees model (GBDT) with LambdaRank to predict a ranked list of transfer languages based on their expected task-score c_{t,a}.
• Construct training data by exhaustively evaluating transfer-language pairs across multiple training task languages to obtain gold-standard rankings.
• Evaluate rankings using Normalized Discounted Cumulative Gain (NDCG@3) and compare LangRank variants (all features, dataset-only, URIEL-only) against baselines.
• Use cross-task experiments (MT, EL, POS, DEP) with leave-one-language-out cross-validation to assess generalization.

实验结果

研究问题

• RQ1 Can a data-driven ranking model improve the selection of transfer languages for cross-lingual NLP tasks over traditional heuristic criteria?
• RQ2 Which features (dataset-dependent vs. dataset-independent) are most informative for predicting effective transfer languages across different NLP tasks?
• RQ3 How does LangRank perform relative to single-feature baselines and to linguistic-distance baselines across MT, EL, POS tagging, and dependency parsing?
• RQ4 What practical guidance on feature importance emerges for informed transfer-language choices?
• RQ5 Are LangRank predictions useful when only typology or dataset features are available (zero-shot or limited-resource settings)?

主要发现

• LangRank 显著优于所有四个 NLP 任务中的单特征启发式方法。
• 将数据集相关特征与语言距离结合通常能够得到最佳的转移语言预测；数据集特征在 MT 和 POS 上尤为占优势。
• 对于 EL，在缺乏完整句子级数据时，某些数据集特征的作用受限，但语言距离仍能提供强有力的指引。
• LangRank（全部特征）通常在 NDCG@3 上优于基线，在某些设置中数据集专用特征对 MT 和 POS 的表现最佳。
• 特征重要性分析揭示了任务特定模式，例如数据集规模和词汇重叠驱动 MT 决策，而地理和句法距离在低数据场景下可能支配 EL 和 DEP。
• 即使只有类型信息可用（URIEL 特征），LangRank 也优于启发式基线，表明在目标任务资源收集前具有实际应用潜力。