TL;DR
This paper introduces a novel computational approach, the "Maximum Entropy Searchlight" model, to determine the authorship era of Old Thai prose texts. By analyzing text segments using varying-length character n-grams, the model provides strong evidence that two historically debated texts, Traiphumikatha and Pumratchatham, were likely written during the Sukhothai era, challenging previous scholarly assumptions.
What Problem Does This Paper Address?
Dating historical texts, especially those in languages with evolving orthography and limited digital resources, is a significant challenge. For Old Thai prose, pinpointing the exact era of authorship for texts like Traiphumikatha and Pumratchatham has been a long-standing debate among literary scholars. Traditional methods rely on qualitative, diachronic close reading, which can be subjective and labor-intensive. Existing Natural Language Processing (NLP) techniques struggle with Old Thai due to a lack of comprehensive corpora and effective tools for word and syllable segmentation, which are drastically different from modern Thai. This paper aims to bridge this gap by developing a computational method that can accurately and interpretably infer authorship eras, even with limited linguistic preprocessing.
Methods
The core of this paper is the Maximum Entropy Searchlight model. This is a multi-class logistic regression model (also known as Maximum Entropy) enhanced with a crucial feature: varying-length character n-grams. Instead of relying on pre-segmented words or syllables, the model operates directly on character sequences of varying lengths (from 2 to 6 characters). This approach circumvents the difficulties of Old Thai segmentation.
To ensure interpretability and manage the high dimensionality of n-gram features, the model incorporates an L1 penalty (Lasso). This penalty effectively shrinks the coefficients of less important features to zero, highlighting the most significant linguistic indicators for each era. The "Searchlight" aspect refers to how the model scans through substrings within text segments, much like a searchlight, to identify distinguishing patterns. The model classifies text into one of three eras: Sukhothai, Ayuddhya, or Rattanakosin.
Datasets
The study utilizes the following datasets:
- Ground Truth Collections:
- Sukhothai era: Stone inscriptions (representing the earliest period).
- Ayuddhya era: Historical Archive on Luang Udomsombat (จดหมายเหตุหลวงอุดมสมบัติ).
- Rattanakosin era: Historical Archive on Luang Udomsombat (จดหมายเหตุหลวงอุดมสมบัติ).
- Texts in Question:
- Traiphumikatha (ไตรภูมิกถา)
- Pumratchatham (ปูมราชธรรม)
The authors used manually cleaned versions of these texts to avoid biases from differing orthography.
Key Insights and Ideas
- Challenging Conventional Wisdom: The model provides computational evidence suggesting that both Traiphumikatha and Pumratchatham were likely written during the Sukhothai era, contradicting the prevailing scholarly belief that Pumratchatham might be from the late Ayuddhya period.
- The Power of Character N-grams: By using varying-length character n-grams, the model effectively bypasses the need for complex and unreliable Old Thai word/syllable segmentation, demonstrating a robust approach for historical text analysis.
- Linguistic Indicators of Eras: The model identifies specific linguistic features, particularly grammaticalized words (words that evolve into grammatical markers) and polyfunctionalized words (words with multiple grammatical roles), as strong indicators of authorship eras. This aligns with linguistic theories of language change.
- Unidirectional Change: The identified linguistic changes, especially those related to grammaticalization, highlight a tendency towards unidirectional change across eras, providing a theoretical framework for understanding language evolution.
- Efficiency of the Model: The model achieves high accuracy with a surprisingly small fraction of the training data, suggesting that stylistic differences between eras are pronounced and readily detectable.
Key Results
- High Accuracy: The model achieves near-perfect accuracy (e.g., 98% micro-average F1 score) on the ground truth collections.
- Sukhothai Era Authorship: The model classifies a significant majority of segments from Traiphumikatha (66.8%) and Pumratchatham (57%) as belonging to the Sukhothai era.
- Minimal Ayuddhya Evidence: For Traiphumikatha, the model found very little evidence supporting an Ayuddhya era authorship (less than 5% of segments).
- Interpretability: The model successfully pinpoints specific words and n-grams that correspond to known linguistic changes studied in Thai historical linguistics, validating its interpretability.
- Data Efficiency: The model requires only about 40% of the text from each era to achieve high accuracy, indicating substantial stylistic differentiation between eras.
- Pumratchatham's Internal Variation: Figure 2 visually demonstrates that the latter half of Pumratchatham shows stylistic shifts that do not align with the Sukhothai era, potentially explaining why its dating has been debated.
Limitations and Future Work
- Limited Scope: The study focuses on only two specific Old Thai prose texts and three historical eras.
- Data Quality: While cleaned versions were used, the inherent quality of historical texts, especially stone inscriptions, can still pose challenges.
- Future Work: The authors suggest that the model's accuracy and interpretability make it suitable for analyzing other Old Thai texts with unknown authorship. Further research could explore applying this model to a wider range of historical texts and linguistic periods, and delve deeper into the specific grammaticalization patterns identified. The model's ability to visualize stylistic shifts within a text also opens avenues for detailed textual analysis.
Summary updated