Jintang Xue, Chengwei Wei, Yun-Cheng Wang, C. -C. Jay Kuo
原文摘要
Static word embeddings give a word one vector, ignoring its different functions as noun, verb or adjective. Training on part-of-speech-tagged text gives each function its own vector, at the cost of a larger table. Using nine controlled corpora, we ask which context words each part of speech needs, whether it needs its own subspace, and whether the number of dimensions a word keeps should follow its part of speech or its frequency. The tags raise verb similarity on SimVerb-3500 at every dimension from 50 to 300 but help nouns much less. Added to verb-only text, adverbs help verbs more than as many random words; adjectives help nouns only slightly. At a fixed dimension, per-part-of-speech PCA bases preserve verb and adjective similarity better than a global basis or frequency-band bases. Nouns need more dimensions than verbs only among frequent words. We therefore allocate dimensions by frequency with these bases: at a third of the original size, only rare words lose dimensions, SimVerb-3500 and SimLex-999 nouns match the uncompressed table, and verb-to-noun retrieval stays within 0.02. Part of speech decides what a word's vector should be; frequency decides how much of it to keep.
Answer accuracy alone does not distinguish failures in acoustic perception, rule knowledge, and rule application. AuDiR (Audio Diagnostic Reasoning) examines these distinctions through controlled input interventions. Each item contains an audio clip, a multiple-choice question, a task-general rule, and a textual description of the acoustic observations needed to answer. Three conditions progressively supply the rule and observations; a fourth removes the audio while retaining the question and rule. The first correct diagnostic condition assigns an operational response label, without identifying a unique internal cause. The current benchmark contains 201 items in 18 families drawn from 9 sources, with answer references documented separately from percept extraction. In a preliminary evaluation of Gemini-3.8-flash, accuracy is 29.3% without supplied information, 31.3% with the rule, and 98.5% with the observations also supplied. No-audio accuracy is 23.9%, with a confidence interval containing the 25.0% chance rate. A family-level knowledge probe associates incomplete rule recall with larger gains from rule provision. These results show the benefit of supplying acoustic observations as text, while incomplete validation, possible recognition shortcuts, and non-monotonic responses limit stronger failure-attribution claims.