Abstract
For the cascaded task of Chinese word segmentation, POS tagging and parsing, the pipeline approach suffers from error propagation while the joint learning approach suffers from inefficient decoding due to the large combined search space. In this paper, we present a novel lattice-based framework in which a Chinese sentence is first segmented into a word lattice, and then a lattice-based POS tagger and a lattice-based parser are used to process the lattice from two different viewpoints: sequential POS tagging and hierarchical tree building. A strategy is designed to exploit the complementary strengths of the tagger and parser, and encourage them to predict agreed structures. Experimental results on Chinese Treebank show that our lattice-based framework significantly improves the accuracy of the three sub-tasks.
Introduction: The Challenge of Chinese Syntax
In many Western languages, word boundaries are clear due to spaces, and part-of-speech (POS) tagging is highly accurate. However, Chinese presents a unique challenge: there are no natural delimiters between words. Traditionally, researchers used a pipeline approach: first segment characters into words, then assign POS tags, and finally build a syntactic parse tree. The fatal flaw here is error propagation. If the segmenter makes a mistake, the POS tagger is compromised, and the parser has no hope of recovery.
While "joint" models exist to solve all three tasks at once, they often face a "search space explosion." Considering every possible segmentation, tag, and tree structure simultaneously is computationally exhausting. This paper seeks a middle ground: a framework that is joint enough to share information across tasks but efficient enough to be practical.
Methodology: The Word Lattice and Dual Viewpoints
The authors propose a three-stage framework centered around a Word Lattice, which acts as a "Goldilocks" search space—broad enough to capture correct words but compact enough for efficiency.
- Lattice Generation: Instead of picking one "best" segmentation, a character-based model generates the top $N$ segmentations. These are compacted into a "word lattice"—a graph where edges represent potential words. This lattice acts as a high-quality search space, with an "oracle" F1 score of over 99.65%.
- Dual Viewpoints: The framework employs two distinct models that look at the lattice differently: a sequential Lattice-based POS Tagger that looks at local linear context, and a hierarchical Lattice-based Parser that builds a tree structure.
- Dual Decomposition (The Agreement Strategy): Usually, these two models might disagree on which path through the lattice is best. The authors use Lagrangian relaxation to iteratively nudge the models toward a consensus. If the tagger and parser disagree on a word or a tag, a "guidance generator" adjusts the weights (Lagrange multipliers) for the next iteration, effectively telling the models to find a common ground.
Key Insights: Leveraging Complementary Strengths
The power of this framework lies in its ability to leverage the best of both worlds. POS taggers are excellent at identifying local patterns, while parsers understand global sentence structure. By forcing them to agree via dual decomposition, the system allows the parser to provide feedback to the segmenter and tagger.
In a standard pipeline, the parser is a passive recipient of errors. In this framework, if a specific segmentation makes it impossible to build a valid tree, the dual decomposition process will penalize that path and encourage the system to find a more linguistically plausible segmentation.
Results and Performance
The lattice-based framework showed significant improvements across the board compared to a standard pipeline on the Chinese Treebank (CTB) Version 5.0:
- Parsing Accuracy: The framework achieved a Parsing F1 score of 83.42, a substantial jump from the pipeline's 81.17.
- POS Tagging: It reached an F1 of 94.40, outperforming the best previous joint inference systems.
- Word Segmentation: Even the segmentation task improved (F1 97.86), proving that knowing the "big picture" (the parse tree) helps identify individual word boundaries.
Conclusion
The authors conclude that their framework effectively bridges the gap between the simplicity of pipelines and the accuracy of joint models. By using word lattices to constrain the search space and dual decomposition to harmonize sequential and hierarchical viewpoints, they have set a new bar for Chinese syntactic analysis. While the iterative nature of the agreement strategy adds some computational overhead, the gains in accuracy and the reduction of error propagation make it a compelling alternative for high-performance NLP.
Summary updated