Linguistic Insights

This scatter plot displays how Japanese SFX cluster based on their versatility (number of distinct English translations, X-axis) and commonness (how widely shared those translations are, Y-axis).

The data forms a distinctive "trumpet" shape.

Most SFX (the "trumpet bell") have relatively few translations (1-20) with mostly unexceptional stats in terms of how many of those translations overlap with translations of other SFX. The variation in commonness appears to align with my hypothesis about SFX: some represent very specific sounds or act as borderline semantically established onomatopoeia, others represent sounds that speakers could choose to ‘imitate’ in several different ways.

A smaller portion of SFX trail off to the lower-mid right (the "trumpet tube"). These tend to be ‘general-purpose’ SFX, with many contextually viable translations. The positioning and relative linearity of the tube reveals that these SFX are not also outliers in terms of how their translations overlap with other Japanese SFX. The trumpet bell data suggests that overall, Japanese and English SFX map a similar auditory space, breaking it down into sounds which both languages can differentiate relatively cleanly, if not necessarily along the same lines. The tube, on the other hand, seems to reveal that Japanese possesses a small set of versatile sounds like ば, さ, and ど which act like auditory primitives or super-categories. English lacks direct equivalents, so translators resort to diverse contextual replacements.

This force-directed network visualizes semantic relationships between Japanese SFX based on the cosine similarity of their English translation sets, represented as a very high-dimensional binary vector. Each node is an SFX; an edge between two nodes means their translations substantially overlap (using a chosen threshold of cosine similarity ≥ 0.5). SFX with strongly overlapping translation sets are pulled together, surfacing semantic clusters and neighborhoods.

Rather than deriving semantic structure from distributional patterns in large corpora (a technique well-known from its results in natural language processing and machine learning), this visualization builds its space from explicitly attested translation relationships. I treat translation not as a proxy for an underlying semantic ground truth but as a relational signal in its own right. Any semantic space is merely one way of structuring meaning among many — there is no privileged language, space, or learning/embedding technique.

Node size reflects translation count — larger nodes have more translations relative to the current visualized set. Edge opacity reflects cosine similarity: darker edges indicate stronger overlap. Edge thickness is inversely related to similarity: the thinnest, darkest lines connect SFX with near-identical translation sets (often two SFX that share a single translation and nothing else), while thicker, more ghosted edges connect nodes with broader but relatively partial overlap.