Context-Vector Word Sense Disambiguation in Translation

Ambiguous source word

Context sentence (click to include/exclude)

Context window size 4 words

Auto-translate confidence threshold 70%

Decision

Chosen translation

—

Status

—

Margin over runner-up

—

How it works

One source word can map to several target-language meanings (a polyseme). A context-aware translator resolves this by embedding every candidate meaning and every surrounding word as a vector in a shared semantic space, then picking the meaning closest to the sentence's actual context.

context vector  c = Σ wᵢ·vᵢ / |Σ wᵢ·vᵢ|   (active words i, weight wᵢ)
similarity      sim(k) = (c · mₖ) / (|c||mₖ|)   for each candidate meaning mₖ
confidence      P(k) = exp(sim(k)/T) / Σⱼ exp(sim(j)/T)   (softmax, T = temperature)
decision        auto-translate if max P(k) ≥ threshold, else flag for human review

Preset buttons — switch the ambiguous source word (bank / spring / bass / light), each with 2–3 candidate meanings.
Context chips — toggle which words of the surrounding sentence actually contribute to the context vector; more on-topic words pull the arrow toward the correct sense.
Context window — real systems only look at a limited span of neighbouring words; shrinking it drops later context words from the vote even if their chip is on.
Confidence threshold — the minimum softmax confidence a real machine-translation pipeline requires before committing to one meaning automatically; below it, the segment is flagged for a human translator instead of silently guessing.

This is a simplified, hand-placed version of the distributional-semantics idea behind word-sense disambiguation in real MT systems (word2vec neighbourhoods, contextual embeddings in transformer encoders): meaning is inferred from the company a word keeps.

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.