Input Representations for Parsing Discourse Representation Structures: Comparing English with Chinese

Chunliu Wang, Rik van Noord, Arianna Bisazza, Johan BosACL 2021 · DOI

Introduction

Semantic parsing—the process of mapping natural language sentences into formal logical representations—is a fundamental challenge in computational linguistics. This paper focuses on Discourse Representation Structures (DRS), a sophisticated formalism capable of capturing complex linguistic phenomena such as quantification, tense, and coreference. While neural semantic parsers have achieved high performance on English DRS parsing, their applicability to languages with significantly different typological features, such as Mandarin Chinese, remains an open question.

Mandarin Chinese presents unique difficulties for standard NLP pipelines: it is an analytic language that lacks explicit word boundaries (no spaces) and utilizes a vast character set compared to the Latin alphabet. This study investigates whether existing English-centric DRS models can be adapted for Chinese and, crucially, which input granularity—characters, words, or subwords—yields the most accurate semantic representations.

Methodology

The researchers employed a sequence-to-sequence (seq2seq) framework based on a bi-directional LSTM architecture. Although Transformers were tested, the Bi-LSTM provided superior performance for this specific task. The core of the investigation involved comparing five distinct input representation types to determine how granularity affects the model's ability to learn semantic structures.

Examples of English and Chinese input representations across different granularities.

As shown in the examples above, the representations ranged from raw character sequences to standard word-level tokenization. For Chinese, word-level tokenization required an external segmenter (Jieba), while English used the Moses tokenizer. The study also evaluated Byte-Pair Encoding (BPE) at various merge levels (1k, 5k, and 10k) to test the effectiveness of subword units.

To address data scarcity, the authors utilized a "fine-tuning" strategy. Models were first pre-trained on a large "silver" dataset of approximately 97,000 documents from the Parallel Meaning Bank (PMB) 3.0.0, which were automatically annotated. These models were then refined on a high-quality "gold" dataset of 8,403 manually corrected English-Chinese document pairs.

Key Results

The models were evaluated using an F-score that measures the overlap of predicted DRS clauses. The results consistently demonstrated the superiority of character-level models over word-level or subword approaches for both languages.

F-scores for English and Chinese DRS parsing across various input types.

As indicated in the results table, English reached a peak F-score of 88.1 using tokenized characters, while Chinese achieved 76.2 with character-level input. Notably, word-level models performed significantly worse in both languages, trailing the best character models by approximately 5 to 6 points. This suggests that neural models are highly effective at learning morphological patterns and word boundaries directly from characters, and that external word segmenters may introduce noise or errors that hinder the parsing process.

Analysis and Discussion

One of the most striking findings is the "Segmentation Paradox" in Chinese. While word segmentation is traditionally viewed as a mandatory preprocessing step for Chinese NLP, this study found that providing explicit word boundaries (Tokenized Characters) offered no significant benefit over raw character sequences. The model appears capable of internalizing the necessary segmentation logic during training.

Qualitative analysis also revealed "Adverbial Friction" between the two languages. English adverbs often translate to Chinese noun phrases (e.g., "slightly" becoming "a little"). Because the DRS concepts used in this study are based on English WordNet synsets, these structural shifts create a mapping bias that makes it harder for the model to achieve high scores in Chinese, even when the underlying semantic meaning is correctly captured.

Conclusion

This research establishes that character-level input representations are the most robust choice for neural DRS parsing in both English and Mandarin Chinese. By bypassing the need for error-prone external tokenizers, character-level models allow the network to learn more flexible and accurate mappings to formal logic. Future work will explore the integration of large-scale pre-trained models like mBART and the development of more language-neutral ontologies to reduce the current English-centric bias in DRS representations.

Summary updated

BibTeX
@inproceedings{wang2021input,
  author = {Chunliu Wang and Rik van Noord and Arianna Bisazza and Johan Bos},
  title = {{Input Representations for Parsing Discourse Representation Structures: Comparing English with Chinese}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
  doi = {10.18653/v1/2021.acl-short.97},
}
Figure