CISA: Chinese Information Structure Analysis for Scientific Writing with Cross-lingual Adversarial Learning

Hen-Hsen Huang, Hsin-Hsi ChenIJCAI 2018 · DOI

TL;DR

This paper introduces CISA, the first system for analyzing the information structure of Chinese scientific writing. It tackles the challenge of limited Chinese training data by using cross-lingual adversarial learning to transfer knowledge from English datasets, enabling effective advice for Chinese scientific writing without requiring labeled Chinese data.

What Problem Does This Paper Address?

Writing effective scientific papers involves more than just correct grammar and syntax; it requires a clear and logical organization of information at the discourse and knowledge levels. While tools exist to help with English scientific writing, there's a significant gap for Chinese scientific writing. Specifically, identifying the "information structure" of sentences (e.g., classifying them as Background, Purpose, Method, Results, or Conclusion) is crucial for providing high-level writing advice. However, there's a lack of labeled Chinese data to train such systems. This paper aims to bridge this gap by developing a system that can analyze Chinese scientific writing structure, even with limited or no labeled Chinese data.

Methods

The core of CISA is a neural network designed for information structure identification. The key innovation is the use of cross-lingual adversarial learning. Here's how it works:

  • Shared Representation: The system uses bilingual word embeddings and Universal POS tags to represent sentences, aiming for a language-neutral representation.
  • Information Structure Classifier: A GRU (Gated Recurrent Unit) layer encodes the sentence representation to classify it into one of five information structure types.
  • Language Discriminator: An adversarial component is introduced. This discriminator tries to distinguish whether a sentence representation comes from English or Chinese.
  • Adversarial Training: The main network is trained to "fool" the discriminator, forcing it to learn representations that are invariant to language. This allows the model, trained primarily on English data, to generalize to Chinese.
  • Sequence Modeling: To further improve performance, sequence modeling (using the Viterbi algorithm) is applied to consider the entire abstract's structure, not just individual sentences.

Datasets

  • Training Data (English): The NTHU Academic Writing Database, containing 597 labeled English abstracts in the EECS domain, was used for training the main information structure identification model.
  • Evaluation Data (Chinese): 100 Chinese abstracts from the EECS domain were manually annotated as ground-truth for evaluation.
  • Language Discriminator Training Data: 30,000 English and Chinese sentence pairs from the UM-Corpus (a large-scale English-Chinese parallel corpus) in the Thesis domain were used to train the language discriminator.

Key Insights and Ideas

  • Cross-lingual Transfer is Viable for Structure Analysis: The paper demonstrates that knowledge learned from labeled English data can be effectively transferred to analyze the information structure of Chinese scientific writing, even without labeled Chinese data.
  • Adversarial Learning Enhances Language Invariance: The adversarial training mechanism significantly improves the model's ability to generalize across languages by forcing it to learn language-neutral representations.
  • Discourse-Level Advice is Possible: By identifying information structure, systems can move beyond simple grammar checks to provide more sophisticated advice on how to organize scientific arguments.
  • Bilingual Embeddings are Key: The use of pre-trained bilingual word embeddings is fundamental to achieving cross-lingual capabilities.

Key Results

  • The cross-lingual MAIN model (trained on English, tested on Chinese) achieved an F-score of 58.17%, which was slightly better than a monolingual model trained on a small Chinese dataset.
  • The MAIN+ADV+SEQ model, which incorporates adversarial training and sequence modeling, achieved a promising F-score of 71.93% on the Chinese data, representing a significant improvement.
  • In cross-lingual sentence retrieval (for knowledge retrieval), the sentence representation achieved an MRR of 0.70.

Limitations and Future Work

  • Limited Chinese Data: While the system works without labeled Chinese data for structure identification, the overall performance is still influenced by the availability and quality of parallel data for training the adversarial component and the initial bilingual embeddings.
  • Future Work: The authors suggest exploring multi-label classification and finer-grained information structure analysis in the future. They also plan to investigate more advanced cross-lingual transfer learning techniques.

Summary updated

BibTeX
@inproceedings{huang2018cisa,
  author = {Hen-Hsen Huang and Hsin-Hsi Chen},
  title = {{CISA: Chinese Information Structure Analysis for Scientific Writing with Cross-lingual Adversarial Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/853},
}
Figure