A Word Labeling Approach to Thai Sentence Boundary Detection and POS Tagging

Nina Zhou, AiTi Aw, Nattadaporn Lertcheva, Xuancong WangCOLING 2016

TL;DR

This paper introduces a novel "word labeling" approach for Thai Sentence Boundary Detection (SBD), treating spaces as regular words rather than solely as sentence terminators. By jointly modeling SBD with Part-of-Speech (POS) tagging using a Factorial Conditional Random Field (FCRF) model, they achieve improved SBD accuracy and efficiency, particularly with a "2-step" joint model that balances performance and speed.

What Problem Does This Paper Address?

Accurately identifying sentence boundaries is crucial for many Natural Language Processing (NLP) tasks. While languages like English often rely on clear punctuation marks (periods, question marks) for sentence demarcation, Thai presents a unique challenge. Thai writing, especially in modern usage, does not consistently use spaces to mark sentence endings. Spaces can appear within sentences for various grammatical reasons (e.g., after interjections, conjunctions, or in numerical expressions), and conversely, sentences can end without a space. This ambiguity makes traditional space-based SBD methods unreliable. The paper aims to overcome this by proposing a more robust approach that considers the context of words rather than solely relying on the presence or absence of spaces.

Methods

The core innovation is a word labeling approach for Thai SBD. Instead of classifying spaces as sentence boundaries (SB) or non-sentence boundaries (nSB), this method treats every word, including spaces, as a potential unit to be labeled. Each word is then assigned either an SB or nSB label.

To further enhance SBD, the authors propose a joint framework for Part-of-Speech (POS) tagging and SBD. This is implemented using Factorial Conditional Random Fields (FCRFs), a type of graphical model capable of handling multiple layers of labels. The FCRF model allows for the simultaneous prediction of both POS tags and sentence boundaries, leveraging the contextual information from both tasks.

A key development is the "2-step" Joint model. Recognizing that a full joint model with all Thai POS tags can be computationally expensive, this approach first predicts a reduced set of 12 "pseudo" POS tags. These pseudo tags are then used to restore the original, more granular POS tags. This strategy significantly improves processing speed while maintaining high accuracy.

Datasets

The research utilizes two primary datasets:

  • ORCHID corpus: This corpus is used for evaluating the word-labeling approach and comparing it with previous methods. The authors note that for their experiments, they ensured nearly all sentences in the ORCHID corpus lacked explicit sentence-ending spaces to better reflect the challenges of modern Thai writing.
  • TaLAPi corpus: This corpus, specifically its news domain, is used for evaluating the joint POS tagging and SBD framework. It is annotated with word segmentation, POS tags, and named entities.

Key Insights and Ideas

  • Spaces are not reliable sentence delimiters in Thai: The paper strongly demonstrates that relying solely on spaces for Thai SBD is fundamentally flawed due to inconsistent usage.
  • Word labeling is a more robust SBD paradigm: Treating SBD as a sequence labeling task on words, rather than a binary classification of spaces, offers a more flexible and accurate solution.
  • Joint modeling of SBD and POS tagging is beneficial: Incorporating POS information as a feature or a joint label significantly improves SBD performance, suggesting a strong linguistic dependency between these tasks.
  • The "2-step" joint model offers a practical trade-off: By simplifying the POS tagging task in an intermediate step, the "2-step" model achieves substantial speed improvements over a full joint model without a significant drop in SBD accuracy. This makes it more practical for real-world applications.
  • Feature engineering matters, but not always more is better: While adding features like Named Entity Recognition (NER) can improve performance, the paper also shows that simply adding more features doesn't always lead to better results, especially when it comes to pseudo-POS tagging performance in the "2-step" model.

Key Results

  • The proposed word-labeling approach on the ORCHID corpus achieved a 91.70% F-score for SBD, outperforming previous methods like POS-trigram (76.98%), Winnow (84.28%), and ME (84.83%).
  • The joint POS tagging and SBD model on the TaLAPi corpus showed significant improvements. For instance, the "1-step" Joint model yielded a 3% increase in recall and a 2% increase in F-score for SBD compared to an isolated SBD model.
  • The "2-step" Joint model demonstrated remarkable efficiency, reducing processing time by more than half compared to the "1-step" Joint model, while achieving comparable SBD F-scores.
  • With enhanced features (including NER), the "2-step" Joint model achieved the best F-score of 68.87% on the TaLAPi corpus, while also saving 50% of the time compared to the Cascade model.

Limitations and Future Work

  • Feature Optimization: The paper notes that while additional features can improve performance, there's a point of diminishing returns, and not all features consistently benefit all aspects of the model (e.g., pseudo-POS tagging). Future work could explore more targeted feature selection.
  • Interplay between tasks: The authors suggest further exploration into how word boundary information, POS tagging, and sentence boundary information mutually influence each other.
  • Broader Language Contexts: While the focus is on Thai, the insights into the limitations of space-based SBD and the benefits of word-level labeling could be relevant for other languages with similar writing conventions.
  • Real-time performance: While the "2-step" model significantly improves speed, further optimization for real-time applications might be explored.

Summary updated

BibTeX
@inproceedings{zhou2016word,
  author = {Nina Zhou and AiTi Aw and Nattadaporn Lertcheva and Xuancong Wang},
  title = {{A Word Labeling Approach to Thai Sentence Boundary Detection and POS Tagging}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2016},
}
Figure