Parsing Chinese Sentences with Grammatical Relations

Weiwei Sun, Yufei Chen, Xiaojun Wan, Meichun LiuComput. Linguistics 2019 · DOI

Abstract

Modern dependency parsing often relies on tree structures, which struggle to capture the complex syntactic nuances of analytic languages like Mandarin Chinese. This paper introduces a framework for representing Chinese sentences as deep dependency graphs, capturing long-distance dependencies and coordination that traditional trees miss. By leveraging a novel graph-merging approach and contextualized embeddings (ELMo), the authors achieve high accuracy in recovering these intricate grammatical relations.

Introduction

The core challenge in Chinese natural language processing (NLP) lies in its lack of morphological cues. Unlike synthetic languages that use case endings or verb conjugations, Mandarin relies heavily on word order and configuration. Standard dependency trees, where each word has exactly one head, are often too restrictive to represent the underlying "deep" syntax. This paper addresses this by moving toward directed graphs that explicitly represent grammatical relations (GR).

The Need for Deep Dependency Graphs

Trees are computationally efficient but linguistically impoverished. They fail to adequately represent phenomena such as coordination, where multiple verbs share a subject, or long-distance dependencies (LDD), where a word's role is determined by a distant constituent.

A deep dependency graph for a Chinese sentence, illustrating multiple heads and long-distance dependencies.

As shown in Figure 1, a single word can have multiple incoming edges, representing various grammatical roles. For instance, the word "文件" (document) serves as the object of multiple verbs and is involved in a relative clause structure. Capturing these overlapping relations is essential for tasks like Semantic Role Labeling (SRL) and machine translation.

Methodology

The authors propose a three-pronged strategy to build and parse these graphs.

Linguistically-Motivated Extraction

The first step involves converting the Chinese TreeBank (CTB), which uses phrase-structure trees, into deep dependency graphs. This process targets the recovery of "hidden" information, such as traces and empty categories, which are often omitted in surface-level representations.

A phrase-structure tree from the Chinese TreeBank, showing the complex hierarchical structure and empty categories used as the source for graph extraction.

Figure 2 illustrates the complexity of the source phrase-structure trees. The extraction algorithm uses rules inspired by Lexical Functional Grammar (LFG) to map these structures onto the graph format seen in Figure 1.

Graph-Based Parsing via "Graph Merging"

Parsing general graphs is computationally difficult. To solve this, the authors use a "divide and conquer" strategy. They decompose a complex graph into three simpler, tree-like subgraphs. They found that over 98% of sentences in their corpus could be represented by merging just three projective trees. Individual models are trained to predict these subgraphs, which are then unified using Lagrangian Relaxation to ensure a consistent, globally optimal output.

Transition-Based Parsing

In addition to the graph-merging approach, the authors implemented a neural transition-based parser. Using a list-based (Covington-style) system, this parser can create non-projective edges and assign multiple heads to a single word, building the graph incrementally. Both architectures utilize Bi-LSTMs and ELMo embeddings to capture deep semantic context.

Key Insights and Results

A significant finding of the study is the role of ELMo embeddings. In Chinese, part-of-speech (POS) tags are notoriously difficult to predict accurately. The authors discovered that ELMo's contextualized representations are so powerful that they can effectively replace gold POS tags, making the parser much more robust for real-world applications where perfect tags are unavailable.

The extraction algorithm proved highly reliable, achieving 99.17% precision against a manually annotated gold standard. In terms of parsing performance, the Graph Merging model achieved a labeled F-score of 86.17 with gold POS tags, and 84.90 in a realistic scenario using ELMo without gold tags.

Conclusion and Future Work

While the proposed framework significantly advances Chinese dependency parsing, long-distance dependencies remain a challenge. Recall for non-local arcs is lower than for local relations. Future research will focus on refining graph decomposition methods and integrating these deep GR graphs into broader semantic representations like Abstract Meaning Representation (AMR).

Summary updated

BibTeX
@article{sun2019parsing,
  author = {Weiwei Sun and Yufei Chen and Xiaojun Wan and Meichun Liu},
  title = {{Parsing Chinese Sentences with Grammatical Relations}},
  journal = {Comput. Linguistics},
  year = {2019},
  doi = {10.1162/coli\_a\_00343},
}
Figure