Translating Headers of Tabular Data: A Pilot Study of Schema Translation

Kunrui Zhu, Yan Gao, Jiaqi Guo, Jian-Guang LouEMNLP 2021 · OpenReview

Introduction

While Neural Machine Translation (NMT) has achieved remarkable success in translating natural language text, its application to structured data remains a significant challenge. A vast amount of global information is stored in tables, yet translating the "schema"—the headers of these tables—is often overlooked. This paper, presented at EMNLP 2021, introduces the task of Schema Translation and proposes CAST (Context-Aware Schema Translation), a model designed to leverage the structural context of tables to improve translation accuracy.

The Challenges of Schema Translation

Translating table headers is fundamentally different from translating plain text due to two primary factors:

  1. Morphological Differences: Headers are rarely complete sentences. They frequently consist of short phrases, domain-specific abbreviations (e.g., "Loc." for Location), and special symbols (e.g., "$").
  2. Contextual Ambiguity: Headers are often polysemous. For instance, the header "Match" could refer to a sporting event or a physical object. Correct translation requires understanding the surrounding "sibling" headers or the actual data values within the columns.
Examples of domain-specific headers and their paraphrased meanings, illustrating the ambiguity and abbreviation challenges.

General-purpose models like Google Translate often fail to capture this structural context, leading to literal or incorrect translations that diminish the data's utility.

Methodology: The CAST Model

To address these challenges, the authors propose CAST. Instead of treating a header as an isolated string, CAST treats the table structure as a directed graph.

Graph Representation and Relational Edges

The model represents the target header, its sibling headers, and sample cell values as nodes in a graph. It defines specific relationship types, such as "sibling header" and "value belong," to explicitly model the dependencies between different table components.

Relational-Aware Transformer

CAST utilizes a modified Transformer encoder capable of processing these graph relations. This allows the model to "attend" to the most relevant context. For example, seeing numerical values in a column helps the model correctly translate a header like "Area" as a mathematical measurement rather than a geographic region.

The CAST architecture features a relational-aware attention layer that processes a relational matrix (R) to incorporate table structure.

Dataset Construction

The authors developed the first major parallel dataset for schema translation, consisting of 3,158 tables and 11,979 headers across English, Chinese, French, German, Spanish, and Japanese. The data was sourced from WikiTableQuestions and the Spider dataset. Notably, professional translators had to manually correct 40% of the initial machine-translated headers, highlighting the inadequacy of current general-purpose NMT tools for this task.

Key Results and Insights

CAST consistently outperformed state-of-the-art baselines like M2M-100 and MBart-50M2M. In English-to-Chinese translation, CAST achieved a 47.7 BLEU score, a significant improvement over the 27.4 achieved by the base M2M-100 model.

Ablation studies confirmed that the model's awareness of table architecture—specifically the relational edges—is the primary driver of its success. Human evaluations also favored CAST, particularly for its ability to handle abbreviations and polysemy correctly.

Conclusion and Future Work

This pilot study demonstrates that incorporating structural context is essential for accurate schema translation. While CAST represents a significant step forward, limitations remain, including a focus on English as the source language and the computational cost of graph-based modeling. Future research may explore zero-shot schema translation for low-resource languages and the simultaneous translation of both headers and cell content.

Summary updated

BibTeX
@inproceedings{zhu2021translating,
  author = {Kunrui Zhu and Yan Gao and Jiaqi Guo and Jian-Guang Lou},
  title = {{Translating Headers of Tabular Data: A Pilot Study of Schema Translation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  url = {https://openreview.net/forum?id=Z7919xoFnr8},
}
Figure