TL;DR
This paper introduces CSTT, a semi-automatic tool for syntactic tagging of Chinese text using dependency grammar. CSTT learns tagging rules from human input, improving automation and consistency over time, and can also generate phrase structure representations.
What Problem Does This Paper Address?
Building large, syntactically annotated corpora (treebanks) is crucial for Natural Language Processing (NLP) tasks like Machine Translation (MT) and linguistic research. However, manual annotation is time-consuming, difficult to keep consistent, and expensive. Existing methods for syntactic tagging often struggle with efficiency and maintaining high accuracy, especially for languages like Chinese with complex structures. The authors aim to address this by developing a tool that leverages human expertise to semi-automatically build these valuable resources.
Methods
The core of CSTT is a context-dependent shift/reduce parsing mechanism adapted for syntactic tagging. It operates on a sequence of part-of-speech (POS) tagged words. The system uses a dependency grammar formalism, defining 44 types of dependency relations for Chinese.
The key innovation is its human-machine cooperative learning approach:
- Initial Manual Tagging: In the beginning, human annotators perform most of the tagging.
- Rule Acquisition: The system observes human actions (shift/reduce operations, phrase structure construction, and dependency relation assignment) and records them along with the current context (top elements of the stack and input string). These observations form new tagging rules.
- Automated Assistance: As the rule base grows, CSTT can suggest actions to the human annotator or even perform actions automatically.
- Consistency Guarantee: By learning from human input and accumulating rules, the system aims to ensure consistent tagging.
The system also has the capability to produce phrase structure trees, which can be mapped to dependency trees and vice-versa.
Datasets
The experiments were conducted on a small corpus of 1300 sentences of daily life Chinese text. This corpus was pre-processed with word segmentation and POS tagging (using a bi-gram model with ~94% accuracy). No new benchmark datasets were introduced.
Key Insights and Ideas
- Dependency Grammar as a Suitable Formalism: The authors argue that dependency grammar is more economical and convenient for corpus tagging than phrase structure grammar because it lacks non-terminal nodes, directly represents word relations, and maps more straightforwardly to case representations.
- Semi-Automatic Learning is Key: The human-in-the-loop approach is central. By learning from human interventions, the system can gradually increase its automation level while maintaining high consistency, which is a significant advantage over purely manual or fully automated methods.
- Context is Crucial for Disambiguation: The system uses context (elements on the stack and input string) to make tagging decisions, which helps resolve some ambiguities that might otherwise require human intervention.
- Convergence of Rules: The paper suggests that the number of rules needed for syntactic tagging converges, implying that a sufficiently large rule base can cover a high percentage of common sentences.
Key Results
- Rule Acquisition: From the first 300 sentences, 1455 rules were acquired. Over the entire 1300-sentence corpus, a total of 6521 rules were obtained.
- Increasing Automation: The automatic tagging ratio improved as more sentences were tagged and the rule base grew. After tagging 1200 sentences, the automatic tagging ratio exceeded 50%.
- Consistency: The paper emphasizes that the human-machine cooperative approach guarantees good consistency in tagging.
- Efficiency: While not explicitly quantified in terms of speedup, the goal is to improve efficiency by increasing automation.
Limitations and Future Work
- Ambiguity Resolution: CSTT can resolve some contextual ambiguities but struggles with contextual-independent ambiguities, requiring human intervention in such cases.
- Automation Level: The current automatic tagging ratio is above 50% after significant training, but the authors hope to raise this above 80% for practical use.
- Future Work: The authors suggest that with further development, CSTT could evolve into an efficient dependency parser. They also believe the underlying principles could be applied to POS tagging and case tagging. The long-term goal is to achieve nearly full automatic tagging with minimal human intervention.
Summary updated