Semantic Drift
A phenomenon in multi-hop translation or language conversion in which meaning progressively deviates from the original. Manifestations include distortion of modality, negation scope, causal relationships, and specialized terminology, making it a key quality concern in multilingual text processing.
Timeline (last 90 days)
Freezing text embeddings prevents text representations from adaptively adjusting to changes in graph structure, making it difficult to capture fine-grained semantic drift.
Autonomous research systems are highly efficient on structured tasks, but are prone to "semantic drift": when search results do not match the preset topic, the system may still force content generation according to the template, causing the title to be disconnected from the body.
Narrative documents (PRDs, design documents, vision documents) are prone to falling out of sync with actual code, producing Semantic Drift.
Edit Locality measures a model's ability to keep unaffected regions unchanged when executing local modification instructions; many diffusion models cause subtle subject distortion or semantic drift due to global resampling.
Proper nouns, technical terms, and culturally loaded words are typically filled in with the most frequently used general expressions during translation, causing specialized context to be diluted.
In multi-hop translation, after continuous transliteration through five or more languages, the core proposition of a sentence is largely preserved, but the mood of modal verbs, the scope of negation, and implied causal relationships undergo unpredictable shifts.
For semantic drift (where field names and types are unchanged but meanings have changed), there are currently almost no mature automated detection methods beyond manually noticing numerical anomalies.
A bug where the specification and implementation are inconsistent is called semantic drift in software engineering, and is difficult to detect through conventional code review.
All Facts (8)
Freezing text embeddings prevents text representations from adaptively adjusting to changes in graph structure, making it difficult to capture fine-grained semantic drift.
50%UnverifiedAutonomous research systems are highly efficient on structured tasks, but are prone to "semantic drift": when search results do not match the preset topic, the system may still force content generation according to the template, causing the title to be disconnected from the body.
50%UnverifiedNarrative documents (PRDs, design documents, vision documents) are prone to falling out of sync with actual code, producing Semantic Drift.
50%UnverifiedEdit Locality measures a model's ability to keep unaffected regions unchanged when executing local modification instructions; many diffusion models cause subtle subject distortion or semantic drift due to global resampling.
50%UnverifiedProper nouns, technical terms, and culturally loaded words are typically filled in with the most frequently used general expressions during translation, causing specialized context to be diluted.
50%UnverifiedIn multi-hop translation, after continuous transliteration through five or more languages, the core proposition of a sentence is largely preserved, but the mood of modal verbs, the scope of negation, and implied causal relationships undergo unpredictable shifts.
50%UnverifiedFor semantic drift (where field names and types are unchanged but meanings have changed), there are currently almost no mature automated detection methods beyond manually noticing numerical anomalies.
50%UnverifiedA bug where the specification and implementation are inconsistent is called semantic drift in software engineering, and is difficult to detect through conventional code review.
50%