[Paper Review] A Chinese Dataset with Negative Full Forms for General Abbreviation Prediction

Yi Zhang, Xu Sun|arXiv (Cornell University)|Dec 18, 2017

Natural Language Processing Techniques19 references4 citations

TL;DR

This paper introduces a novel Chinese dataset for general abbreviation prediction that includes negative full forms (NFFs)—terms without valid abbreviations—addressing a critical gap in existing corpora. The dataset, built from reliable NLP resources and preprocessed with word segmentation and POS tagging, enables training and evaluation of models like CRF, MEMM, perceptron, and bidirectional LSTM, with the latter achieving the best performance on both accuracy and NFF discrimination tasks.

ABSTRACT

Abbreviation is a common phenomenon across languages, especially in Chinese. In most cases, if an expression can be abbreviated, its abbreviation is used more often than its fully expanded forms, since people tend to convey information in a most concise way. For various language processing tasks, abbreviation is an obstacle to improving the performance, as the textual form of an abbreviation does not express useful information, unless it's expanded to the full form. Abbreviation prediction means associating the fully expanded forms with their abbreviations. However, due to the deficiency in the abbreviation corpora, such a task is limited in current studies, especially considering general abbreviation prediction should also include those full form expressions that do not have valid abbreviations, namely the negative full forms (NFFs). Corpora incorporating negative full forms for general abbreviation prediction are few in number. In order to promote the research in this area, we build a dataset for general Chinese abbreviation prediction, which needs a few preprocessing steps, and evaluate several different models on the built dataset. The dataset is available at https://github.com/lancopku/Chinese-abbreviation-dataset

Motivation & Objective

• To address the lack of datasets that include negative full forms (NFFs) in Chinese abbreviation prediction, which are common in real-world text but underrepresented in existing corpora.
• To build a comprehensive, reliable dataset of Chinese full forms—both those with valid abbreviations and those without—to support general abbreviation prediction research.
• To evaluate multiple machine learning models, including CRF, MEMM, averaged perceptron, and bidirectional LSTM, on this new dataset to assess performance in both positive and negative full form prediction.
• To demonstrate that neural networks, particularly bidirectional LSTM, outperform traditional models in handling the complexity of general abbreviation prediction, including distinguishing NFFs.

Proposed method

• The dataset is constructed from established Chinese NLP corpora, such as People's Daily, and includes both positive full forms (with valid abbreviations) and negative full forms (no valid abbreviations).
• Preprocessing steps include word segmentation and part-of-speech (POS) tagging to provide linguistic features for model input.
• Each character in a full form is assigned a segmentation tag and POS tag, which are embedded into 20-dimensional vectors and concatenated with 50-dimensional character embeddings to form the input for sequence models.
• Bidirectional LSTM (BLSTM) is employed to capture both past and future context, with a hidden size of 200 (100 forward, 100 backward), and the model predicts abbreviation characters via a conditional labeling mechanism.
• The CRF model is used as a sequence labeling baseline, leveraging global feature dependencies, while MEMM and averaged perceptron serve as traditional discriminative models for comparison.
• Model evaluation uses two metrics: All-match accuracy (comparing full predicted abbreviations to gold standard) and Character accuracy (per-character prediction accuracy).

Experimental results

Research questions

• RQ1 Can a large-scale Chinese dataset that includes negative full forms (NFFs) improve the performance of general abbreviation prediction models?
• RQ2 How do traditional sequence models (CRF, MEMM, averaged perceptron) compare to deep learning models (BLSTM) in predicting abbreviations, especially in distinguishing NFFs?
• RQ3 To what extent does incorporating part-of-speech and segmentation information enhance abbreviation prediction performance?
• RQ4 Does a bidirectional LSTM outperform other models in capturing contextual dependencies for abbreviation generation, particularly in handling irregular abbreviation patterns in Chinese?

Key findings

• The bidirectional LSTM model achieved the highest overall performance, outperforming CRF, MEMM, and averaged perceptron in both All-match accuracy and character-level accuracy.
• The CRF model showed strong performance, particularly in character-level accuracy, indicating its effectiveness in modeling sequential dependencies for abbreviation prediction.
• The simple heuristic baseline—using the first character of each word—performed poorly, highlighting the need for sophisticated models to handle irregular Chinese abbreviation patterns.
• The inclusion of negative full forms (NFFs) significantly increases the difficulty of the task, as evidenced by lower performance across all models, especially in discriminating NFFs from positive full forms.
• The dataset, available at https://github.com/lancopku/Chinese-abbreviation-dataset, provides a valuable benchmark for future research in Chinese abbreviation prediction with NFFs.