An Exploratory Analysis on the Explanatory Potential of Embedding-Based Measures of Semantic Transparency for Malay Word Recognition

TL;DR

The paper addresses how semantic transparency can be captured for Malay word recognition using embedding-based measures. It combines t-SNE visualization of semantic geometry with five embedding-derived metrics, including LDA-based prefix predictability and centroid/FRACSS correlations, to predict lexical decision latencies. In behavioral analyses, all measures significantly predict reading times beyond traditional predictors, with centroid-to-prefix similarity delivering the best fit. The work advances realizational morphology in Malay and provides data and measures for the Malay Lexicon Project.

Abstract

Studies of morphological processing have shown that semantic transparency is crucial for word recognition. Its computational operationalization is still under discussion. Our primary objectives are to explore embedding-based measures of semantic transparency, and assess their impact on reading. First, we explored the geometry of complex words in semantic space. To do so, we conducted a t-distributed Stochastic Neighbor Embedding clustering analysis on 4,226 Malay prefixed words. Several clusters were observed for complex words varied by their prefix class. Then, we derived five simple measures, and investigated whether they were significant predictors of lexical decision latencies. Two sets of Linear Discriminant Analyses were run in which the prefix of a word is predicted from either word embeddings or shift vectors (i.e., a vector subtraction of the base word from the derived word). The accuracy with which the model predicts the prefix of a word indicates the degree of transparency of the prefix. Three further measures were obtained by comparing embeddings between each word and all other words containing the same prefix (i.e., centroid), between each word and the shift from their base word, and between each word and the predicted word of the Functional Representations of Affixes in Compositional Semantic Space model. In a series of Generalized Additive Mixed Models, all measures predicted decision latencies after accounting for word frequency, word length, and morphological family size. The model that included the correlation between each word and their centroid as a predictor provided the best fit to the data.

Figures (5)

• Figure 1: Note. Each coloured dot represents a derived word and each colour corresponds to a particular prefix a word contains. Words that have similar embeddings appear closer to each other in the t-SNE space than others.
• Figure 2: Note. Vector illustration. Solid lines indicate vectors for the base and derived word, and the dotted line represent the shift vectors. Of interest is the angle between the base and derived vectors, and the derived and shift vectors from the point of origin at (0, 0).
• Figure 3: Note. Top row: Interaction between frequency and root family size (left). Data represented by dark blue points. Warmer colours (e.g., pink, orange) on the left-hand side denote longer RTs and cooler colours (e.g., green) on the right-hand side denote shorter RTs. Numbers on contour lines represent fitted inverse RT values. Partial effect of word length (right), with rugged lines on the x-axis representing the distribution of the data. Bottom row: Partial effects of trial number (centered and scaled) by subjects (left)
• Figure 4: Note. Partial effect of each correlation measure on RT. Black lines on the x-axis of each plot represent the data
• Figure 5: Note. Correlation heatmap of prefix centroid embeddings and a subset of the linear comprehension mapping F that maps the embeddings of a word form (trigrams) to its meaning (FastText) in the DLM. The set of trigrams presented correspond to at least one of the prefixes. Darker shades, compared to lighter shades, indicate a stronger correlation between the embeddings of the centroid and comprehension mapping F. The strongest correlations are present for the trigrams that correspond to the prefix, indicating that it is the prefixal trigrams that contribute most to realizing the meaning of the centroid