Abstract
Disambiguating abbreviations in specialized domains like Jewish Law (Halakhic) literature is challenging due to high ambiguity and the absence of explicit definitions. This study proposes a machine-learning-based approach that combines variants of the "One Sense Per Discourse" (OS) hypothesis with Support Vector Machines (SVM). By utilizing 18 distinct features—ranging from statistical to contextual—without relying on complex natural language processing (NLP) tools, the system achieves a peak accuracy of 96.09%. The results demonstrate that consistency across a writer's body of work (One Sense Per Writer) and local context are powerful indicators for disambiguation.
Introduction
Abbreviations in Hebrew religious texts are notoriously ambiguous, often having up to eight possible meanings. Unlike modern technical writing, these ancient and dense documents rarely define abbreviations, assuming reader expertise. Traditional disambiguation methods often fail in this context because the "long form" of an abbreviation rarely appears. This research aims to develop a language-independent, "NLP-lite" system to solve these ambiguities without requiring built-in dictionaries or grammar taggers.
Methods
The researchers defined 18 features across three categories: statistical (e.g., most common sense), Hebrew-specific (e.g., Gimatria), and contextual (surrounding words). They applied the One Sense Per Discourse (OS) hypothesis at three levels: Sentence (osS), Article (osA), and Writer (osW). The "Combined" OS form (osSC, osAC, osWC) was also tested, which uses local context first and falls back to discourse-level majorities only if the primary method fails.
The study utilized a specialized dataset of Jewish Law documents written by two prominent scholars, Rabbi Y. M. HaCohen and Rabbi O. Yosef, totaling over 564,000 words. The dataset contained 114,814 abbreviations, of which 42,687 were ambiguous. Four machine learning algorithms were evaluated: Artificial Neural Networks (ANN), Naïve Bayes (NB), J48 (Decision Trees), and Support Vector Machines (SVM).
Results and Key Insights
A significant finding was that contextual features looking after an abbreviation were more predictive than those looking before. For instance, the A1WC feature (one word after) reached 87.75% accuracy under the osWC variant. The "Writer Common Rule" (WC) was the strongest standalone feature at 82.84%, suggesting that scholars are remarkably consistent in their abbreviation usage across their entire body of work.
As shown in the incremental analysis, the SVM model reached a peak accuracy of 96.09% when combining all 18 features. The use of OS variants provided an average relative improvement of approximately 18% over standard feature forms. The "Combined" OS form outperformed "Pure" OS by accounting for rare cases where an author might use different senses of the same abbreviation within a single text.
Conclusion
The high accuracy achieved without formal grammar parsers underscores the effectiveness of statistical and contextual patterns in specialized domains. This "NLP-lite" approach ensures high portability to other languages or domains. Future work will involve testing this model on standard "bag-of-words" representations and exploring the integration of part-of-speech tags to further enhance accuracy.
Summary updated