Technical Summary: Lost in Translation? Exploring the Shift in Grammatical Gender from Latin to Occitan

1. Problem Statement

The paper addresses the diachronic evolution of grammatical gender from Vulgar Latin to Medieval Occitan, specifically focusing on the transition from a tripartite system (masculine, feminine, neuter) to a bipartite one (masculine, feminine). While historical linguistics establishes that Latin neuter nouns were largely absorbed into the masculine gender, the specific factors governing this reassignment in Occitan—whether semantic, phonological, or morphological—remain insufficiently studied.

The study operates within the constraints of a low-resource historical setting. Medieval Occitan is characterized by:

• Data Scarcity: Limited annotated corpora.
• Orthographic Instability: Extensive spelling variations within and across texts due to the lack of a standardized norm.
• Dialectal Fragmentation: Absence of a unified standard.

The core research questions are:

1. RQ1 (Lexical): To what extent can the grammatical gender of Occitan nouns derived from the Latin neuter class be predicted solely by word-level features (phonological and morphological)?
2. RQ2 (Contextual): How is gender information distributed between lexical features and sentence-level context, and how do these sources contribute to model predictions?

2. Methodology

Data and Preprocessing

The study utilizes three annotated sources: Lo Codi (law), Chanson de la Croisade Albigeoise (poetry), and the DOM Dictionary Project. The dataset comprises Latin–Occitan lemma pairs with grammatical gender labels.

• Tokenization Strategy: Recognizing that standard tokenizers fail with historical orthographic noise, the authors propose a hybrid tokenization policy. This combines Occitan-adapted Byte-Pair Encoding (BPE) with a word-level fallback. This approach achieves zero Out-of-Vocabulary (OOV) rates and superior masked token recovery compared to standard WordPiece or subword-only segmentation.
• Embedding Selection: A comparative study of FastText, mBERT, and ByT5 identified mBERT as the most effective backbone for capturing Medieval Occitan variation and cross-lingual structure.
• Domain Adaptation: The mBERT model underwent domain-adaptive Masked Language Modeling (MLM) fine-tuning on raw Occitan texts to better fit the corpus.

Experimental Framework

The authors employ an interpretable deep learning framework to disentangle lemma-internal signals from contextual signals.

A. Lexical-Level Prediction (RQ1)

• Input: Isolated Occitan lemmas and their corresponding Latin lemmas.
• Features:
  • Morphological/Phonotactic: Initial substrings, suffix character n-grams (1–4), syllable counts, and VC (Vowel-Consonant) templates.
  • Stress Proxy: A heuristic approximation of stress position (ultimate/penultimate/antepenultimate).
  • Embeddings: Frozen representations from mBERT, ByT5, and FastText.
• Models: A diverse set of classifiers ranging from linear models (Logistic Regression) to non-linear ensembles (Random Forest, XGBoost) and neural architectures (FFN, BiLSTM, Attention-based models).
• Validation: Lemma-grouped 10-fold cross-validation to prevent data leakage between orthographic variants of the same lemma.

B. Contextual Prediction (RQ2)

• Input: Occitan sentences containing the target noun, linked to Latin lemmas.
• Configurations:
  1. Word-only: Uses isolated embeddings and Latin metadata.
  2. Context-focused: Uses noun-conditioned attention over the full sentence encoding.
  3. Masked-context: The target noun is replaced with [MASK], forcing the model to predict gender based only on surrounding agreement cues (articles, adjectives).
• Analysis: The study uses SHAP values for feature attribution and PoS-conditioned occlusion to quantify the contribution of specific parts of speech (e.g., determiners, adjectives) to the prediction.

3. Key Results

Lexical-Level Performance

• Model Performance: Neural sequence models, particularly a 2× BiLSTM with Multi-Head Self-Attention (MHSA), outperformed shallow baselines. The best configuration used mBERT embeddings with focal loss and class weighting to handle imbalance.
• Feature Importance: Ablation studies revealed that suffix character n-grams (especially Latin and Occitan endings) are the most predictive features, causing the largest performance drops when removed (1.6–1.8 Macro-F1 points). Length and meta-features were secondary contributors. Stress proxies had minimal impact.
• Baseline: The best lexical-only model achieved a Macro-F1 of 0.8224 (mBERT).

Contextual Influence

• Performance Gain: Incorporating sentence-level context significantly improved performance. The Context-focused model achieved a Macro-F1 of 0.929, a substantial increase over the word-only baseline (0.665).
• Masked Context: Even when the noun was masked, the model achieved a Macro-F1 of 0.902, demonstrating that surrounding agreement cues (articles, adjectives) carry strong gender information.
• Disambiguation: In cases where the lemma representation was ambiguous (e.g., ome), the contextual model successfully shifted predictions to the correct gender by leveraging agreement tokens (e.g., the article aquill).
• Attribution: SHAP and occlusion analyses confirmed that nouns, determiners, and adjectives provide the most reliable positive contextual evidence, while conjunctions and prepositions contributed negative or negligible signals.

Gender Shift Patterns

• The data confirms the absorption of Latin neuters into Occitan masculine (3,055 cases) and feminine (1,448 cases).
• While the ending -um is strongly associated with masculine outcomes, it is paradoxically the most common ending for nouns that shifted to feminine, highlighting that raw frequency counts are insufficient to explain the shift without considering morphological cues.

4. Significance and Contributions

The paper claims the following contributions to the fields of NLP and historical linguistics:

1. Interpretable Framework for Low-Resource History: It introduces a general framework for disentangling lemma-internal and contextual signals in grammatical gender prediction, specifically tailored for noisy, low-resource historical languages.
2. Tokenization Innovation: It demonstrates that standard tokenization strategies are insufficient for Medieval Occitan and proposes a hybrid tokenization policy that improves robustness against orthographic variation and spelling noise.
3. Quantifying Gender Distribution: The study characterizes the distribution of gender information, showing that while morphological suffixes provide the primary structural encoding, sentence-level agreement (articles, adjectives) is critical for disambiguation and resolving lexical ambiguity.
4. Resource Availability: The authors release their codebase, datasets, and results to facilitate further research in historical NLP.

5. Limitations

The authors modestly acknowledge several constraints:

• Data Imbalance: The dataset remains relatively small with a skewed gender distribution (approx. 2:1 masculine-to-feminine), which may limit generalization for the minority class despite mitigation techniques.
• Heuristic Preprocessing: Key components, such as the fuzzy-matching threshold and stress-position proxy, were set heuristically and may introduce noise.
• Tagging Accuracy: Part-of-Speech (PoS) attribution relies on automatic tagging with ~71% accuracy, potentially biasing the attribution results.
• Scope: The findings pertain to the synchronic prediction of gender in Medieval Occitan and the Latin-to-Occitan neuter collapse. The paper does not claim to definitively solve the diachronic question of what drove the historical reassignment, noting that this would require different experimental designs and parallel diachronic data.