TL;DR
This paper introduces a two-stage approach for automatically guessing the part-of-speech (POS) of unknown Chinese words. It first uses internal word structure features with CRFs for initial guessing and then refines low-confidence predictions using global context. This method significantly improves precision over prior work, especially for challenging disyllabic words.
What Problem Does This Paper Address?
Chinese, like many East Asian languages, doesn't use spaces to delineate words, making word segmentation a crucial first step in natural language processing (NLP). Building a comprehensive dictionary is impossible due to the constant creation of new words. These "unknown words" can significantly hinder downstream NLP tasks. While detecting unknown words is relatively mature, accurately guessing their POS is a persistent challenge. Traditional POS tagging relies on lexicons, which are unavailable for unknown words. This paper tackles the problem of assigning the correct POS tag to these unknown words, aiming to improve the robustness of Chinese NLP systems.
Methods
The proposed method employs a two-stage approach:
- Internal Component Feature Model (Stage 1): This model treats POS guessing as a sequence labeling problem. It utilizes Conditional Random Fields (CRFs), a powerful framework for sequence modeling, to predict the POS of unknown words based on their internal structure. The features considered include the characters themselves, their lengths, and their POS tags (if known from segmentation). The authors explore different segmentation schemes (character-based vs. immediate constituent-based) and feature combinations, finding that character-based segmentation performs better for derivative words and immediate constituent-based for compounds.
- Global Contextual Feature Model (Stage 2): For words where the first model's prediction has low confidence, a second model is employed. This model leverages global contextual information. It uses a "voting by syntactical templates" approach, inspired by linguistic observations. For a given unknown word, phrases are constructed based on predefined templates (e.g., "不 + word," "开始 + word"). These phrases are then used to query a search engine (Baidu.com) to retrieve snippets. The snippets act as "voters" to assign scores for the word being a noun, verb, or adjective. A threshold is used to determine if the voting result is reliable enough to revise the initial prediction.
A crucial element linking these two stages is a credibility computation mechanism. After the first model makes a prediction, a credibility score is assigned based on the sequence type of the word's internal components. Words with scores below a certain threshold are then passed to the second stage for revision.
Datasets
- Training Data: The Modern Chinese Grammar Information Dictionary (Yu, 1998) was used. This dictionary was segmented and POS-tagged using ICTCLAS 1.0. Monosyllabic words were excluded.
- Testing Data: The Contemporary Chinese Corpus of Peking University (Yu et al., 2002) was used. This corpus, consisting of news articles from the People's Daily, was also segmented and POS-tagged. Unknown words were identified as those present in the corpus but not in the training dictionary.
Key Insights and Ideas
- Internal Structure is Powerful: The paper strongly advocates for the effectiveness of internal component features for POS guessing, demonstrating that a model based solely on these features can achieve high performance.
- Context as a Refinement: Global context is not a replacement but a valuable complement to internal features, particularly for ambiguous words. The two-stage approach effectively leverages both.
- Credibility as a Bridge: The concept of assigning a credibility score to initial predictions is key to intelligently deciding when to apply the more computationally intensive contextual model. This prevents unnecessary processing and focuses resources on problematic cases.
- Linguistic Templates for Context: The use of syntactical templates derived from linguistic rules to extract contextual features is an elegant way to capture semantic and syntactic roles of words in sentences.
- Search Engines as Corpora: Utilizing search engines to gather contextual examples for unknown words is a practical and scalable approach, especially given the ephemeral nature of some new words.
Key Results
- Stage 1 Performance: The internal component feature model (specifically Scheme 1, character-based segmentation) achieved a precision of 93.40% for all words and 86.60% for disyllabic words. This is a significant improvement over previous state-of-the-art results (89% and 74%, respectively).
- Stage 2 Improvement: The second model, using global context, further improved the overall precision by 0.80% for all words and 1.30% for disyllabic words.
- Specific Word Types: The character-based scheme (Scheme 1) was found to be better for derivative words, while the immediate constituent-based scheme (Scheme 3) was better for compounds.
- Reduplication Rules: Two specific rules for reduplicated words improved the precision for four-character words to 97.10%.
Limitations and Future Work
- Training Data Quality: The authors acknowledge that the training data, generated by automatic segmentation and POS tagging tools (like ICTCLAS), contains errors. These errors, which were not revised in the experiment, likely impacted performance.
- Manual Revision of Training Data: A proposed future improvement is to manually revise the training data to correct these errors, which is expected to further boost results.
- Applicability to Other Languages: While the experiments focused on Chinese, the authors believe the method is applicable to other Asian languages, such as Japanese.
- Incorporating More Rules: While the CRF model implicitly learns many rules, explicitly incorporating more linguistic rules could further simplify the machine learning approach.
Summary updated