TL;DR
This paper presents Peking University's contribution to the First International Chinese Word Segmentation Bakeoff, focusing on their role as both a corpus provider and a participant. They argue that an absolute definition of word boundary isn't necessary, and differing segmentation results can be acceptable if they lead to correct syntactic analysis.
What Problem Does This Paper Address?
Word segmentation is a foundational step in Chinese information processing, but achieving consistent and accurate segmentation is a significant challenge. Different segmentation approaches can lead to varying word boundaries, which in turn impacts downstream tasks. The paper addresses the difficulty of establishing a universally accepted evaluation method for word segmentation and explores the implications of different "word" definitions, particularly between Mainland China and Taiwan, on segmentation performance.
Methods
The authors describe their work in two capacities:
- Corpus Provider: They provided a training and test set derived from People's Daily, chosen for its standard Chinese, minimal errors, and broad coverage of linguistic phenomena. This corpus was manually validated for high correctness and consistency. They also emphasize the importance of a detailed specification for corpus processing and NLP evaluation to ensure a fair contest.
- Participant: Their segmentation system was trained on People's Daily data. A key challenge they addressed was the difference in word definitions and lexicon between Mainland China (their primary focus) and Taiwan, as represented by the Penn Chinese treebank. They adapted their lexicon to include Taiwanese terms, though they acknowledge limitations due to unfamiliarity and time constraints. They also mention using rule-based methods to handle certain word patterns.
Datasets
- Training and Test Corpus: Provided by Peking University, derived from People's Daily (January, 1998).
- Evaluation Corpus: The Penn Chinese treebank was used for testing their participant system.
Key Insights and Ideas
- No Absolute Word Boundary Needed: The authors propose that an absolute, universally agreed-upon definition of word boundary for all Chinese segmenters is not essential. Different segmentation outcomes can be valid if they ultimately contribute to correct syntactic analysis.
- Lexical Differences Matter: Significant differences exist in word usage and lexicon between Mainland China and Taiwan. This poses a challenge for segmenters trained on one region's data when evaluated on another, requiring careful adaptation.
- Evaluation is Difficult: The paper highlights that developing a scientific and appropriate evaluation method for word segmentation is as challenging, if not more so, than the segmentation task itself.
- Downstream Task Dependency: While word segmentation is critical, a satisfactory result doesn't guarantee success in subsequent information processing steps.
Key Results
The paper doesn't present specific quantitative accuracy scores for their participant system. However, they state that their segmenter "failed to score when unknown words are involved." They also note that their provided corpus was highly regarded and elected by other participants. Qualitatively, they show examples of how their system's segmentation aligns with or differs from the Penn Chinese treebank's segmentation, illustrating the lexical and definitional discrepancies they encountered.
Limitations and Future Work
- Unknown Words: The recognition of unknown words remains a significant bottleneck for their word segmentation technique, and they acknowledge not having found a good solution yet.
- Taiwanese Lexicon Adaptation: Due to time limitations and unfamiliarity with Taiwanese Chinese, their adaptation of the lexicon to include Taiwanese terms was not exhaustive, leading to potential performance issues.
- Small Training Corpus: The training corpus for their participant system was small, which limited their ability to fully mitigate the negative effects of lexical differences.
- Future Work: They express confidence in making progress on unknown word recognition in the future and emphasize the need for continued fundamental research into the Chinese language for improved automation. They also suggest that future processing programs should be designed to tolerate or correct errors from previous steps.
Summary updated