TL;DR
This paper tackles the challenge of cross-linguistically mapping how languages express "when" clauses, particularly focusing on morphologically encoded temporal subordination common in Latin America and the Caribbean. By developing a novel subtoken-based approach using character n-grams and probabilistic semantic maps, the authors can identify and visualize these complex grammatical patterns, offering a more comprehensive understanding of temporal subordination beyond purely lexical markers.
What Problem Does This Paper Address?
Understanding how languages express temporal relationships, specifically "when" clauses, is crucial for linguistic typology. While English primarily uses a single, underspecified conjunction ("when"), other languages can have multiple lexical options or, more commonly in regions like Latin America and the Caribbean, rely heavily on morphological marking on the verb itself. This morphological variation is notoriously difficult to study using traditional methods that focus on whole words or lexemes, leading to a gap in our understanding of how these languages structure temporal subordination. Existing approaches often struggle to capture this morphological complexity, leading to a bias towards languages with simpler, lexically driven systems. The paper aims to bridge this gap by developing a method that can identify and map both lexical and morphological expressions of "when" clauses across a diverse set of languages.
Methods
The core of this paper's innovation lies in its "subtoken-based" approach, specifically using character n-grams. Instead of relying solely on full word alignments, the authors extract sequences of characters (n-grams) from the English "when" clauses and their corresponding translations. These n-grams are then used to identify potential morphological markers.
Here's a breakdown of the key techniques:
- N-gram Search: Character n-grams (sequences of 2-9 characters) are extracted from the English "when" and aligned tokens in the target languages. Association measures (specifically, chi-squared, χ²) are used to identify n-grams that are highly correlated with "when" clauses. This helps pinpoint potential morphological morphemes.
- Clustering: The identified n-grams are clustered using DBSCAN to group together potential allomorphs of the same morpheme.
- Probabilistic Semantic Maps: The core visualization technique is the creation of probabilistic semantic maps. These maps use Multidimensional Scaling (MDS) to represent the semantic space of "when" clauses. Each point on the map represents a specific instance of a "when" clause, and the distance between points reflects their semantic dissimilarity.
- Geostatistical Interpolation (Ordinary Kriging): To create smooth contour maps of semantic areas, Ordinary Kriging is employed. This technique interpolates the linguistic data (lexical items or n-gram placeholders) across the semantic space, allowing for the visualization of the spatial intensity of different linguistic means. This is a novel application for linguistic typology in this context.
- Benchmarking with Huichol: The Huichol language, with its well-documented switch-reference system and a lexified "when" subordinator, serves as a crucial benchmark. This allows for the calibration of hyperparameters and validation of the n-gram detection method against known grammatical structures.
Datasets
The primary dataset used is a subset of the Mayer and Cysouw's (2014) massively parallel corpus of New Testament translations.
- Latin American and Caribbean Subset: The authors manually created a GeoJSON dataset to define the geographical region of interest and filtered the corpus to include approximately 335 language varieties from Latin America and the Caribbean. This subset is crucial for focusing on languages where morphological subordination is prevalent.
- Huichol Annotation: For benchmarking, the Huichol translation of the New Testament was further annotated with specific switch-reference markers (DS and SS) based on existing typological descriptions. This enriched dataset allowed for a direct comparison between automatically detected n-grams and known morphological markers.
Key Insights and Ideas
- Morphological "When" is Mappable: The most significant insight is that complex morphological strategies for expressing "when" clauses, particularly switch-reference marking, can be effectively detected and mapped using a subtoken-based approach. This opens up the study of morphologically rich languages to large-scale typological analysis.
- Semantic Space of "When" is Diverse: The semantic maps reveal that the concept of "when" is not monolithic. Different languages carve up this semantic space differently, using distinct lexical items or morphological strategies to express nuances of temporal subordination.
- Huichol as a Microcosm: The Huichol benchmark demonstrates the power of the method. The semantic map clearly distinguishes between the lexified subordinator ("quepaucua") and the morphological switch-reference markers (DS and SS), showing how these different strategies occupy distinct regions in the semantic space.
- Beyond Switch-Reference: While switch-reference is a prominent example, the n-gram approach is generalizable and can potentially identify other types of morphological subordination beyond this specific system.
- Geostatistical Mapping Adds Nuance: The use of Kriging allows for a more nuanced understanding of spatial distribution. It can reveal areas where a particular linguistic means is concentrated, even if it's not the most frequent overall, suggesting its importance in defining that semantic region.
Key Results
- High Accuracy for Huichol: The automatic word-alignment combined with the n-gram search method achieved a precision of 0.90, recall of 0.99, and F1-score of 0.94 when benchmarked against manually annotated Huichol data. This indicates a high degree of success in identifying morphological markers.
- Visualization of Linguistic Variation: The paper presents numerous Kriging maps for various Latin American languages, visually demonstrating the diverse ways "when" clauses are encoded. These maps show clear clusters for lexified subordinators, n-gram groups (likely morphological markers), and areas of "NOMATCH" (indicating absence of a direct parallel).
- Identification of Morphological Patterns: The maps successfully identified n-gram groups corresponding to switch-reference markers in languages like Huichol and Quechuan varieties, validating the method's ability to capture these specific morphological phenomena.
- Discovery of Inverted Patterns: The study also identified languages with an "inverted pattern," where lexified means occupy the top of the map (typically associated with morphological means) and n-gram areas are at the bottom, suggesting further complexity and variation in how semantic space is divided.
Limitations and Future Work
- Huichol Bias: The primary limitation acknowledged is that the hyperparameter optimization for the Kriging models was based on a single language (Huichol). This might introduce a bias towards languages with similar subordination systems (e.g., a lexified counterpart to "when" alongside switch-reference morphology), potentially obscuring other typological dimensions.
- N-gram Representation: The n-gram approach identifies groups of characters but doesn't directly provide a single "representative morpheme." The paper notes that groups are numbered, and further work would be needed to select a canonical form from a cluster.
- Limited Scope of Morphological Means: Switch-reference is only one type of morphological subordination. The study acknowledges that other systems exist and might not be fully captured by this specific approach.
- Future Work: The authors suggest several avenues for future research:
- Experimenting with different n-gram sizes and association measures.
- Using other languages as benchmarks for calibration.
- Conducting close-reading evaluations of languages with opposite patterns to Huichol to better understand the method's performance.
- Systematically analyzing the semantic dimensions identified in the maps to discover cross-linguistically relevant "gram types."
- Developing methods to select representative morphemes from n-gram clusters.
Summary updated