Fig 1.
Experimental design and trials structure.
Participants named colored objects in three ways, depending on task instruction: as phrases (white lamp), as single nouns (lamp), as single adjectives (white) or as adjective-noun lists (green, lamp), naming the color of the background followed by the object name. Our analyses assessed the decodability of adjective and/or noun representations in these four contexts.
Fig 2.
Explanation of data, model, and testing procedures.
Values that are fixed appear as solid-colored rectangles, and values that are learned or predicted appear as dotted rectangles. A) The dimensions of the MEG data X (blue) and the word embedding matrix Y (green). B) The process for predicting one dimension (j) of the word embedding matrix Y. Note that this is corresponds to hj(X) in the in-text equations. C) Predicting all dimensions of a word embedding for MEG data sample xi. W is the concatenation of w vectors from B). D) The 2 vs 2 test. The 2 vs. 2 test measures how similar the predictions (,
) are to their corresponding ground truth vectors (ya, yb) using a vector distance criterion d(v,u). If the correct matching of true to predicted vectors (blue lines) represents a smaller distance than the incorrect matching (red lines), the 2 vs 2 test passes.
Fig 3.
Decodability across time for adjectives and nouns when presented within phrases, lists or in isolation as single words.
The grey shading indicates a significant main effect of category on decodability across all contexts, with nouns showing higher accuracy than adjectives in the mid-latency time-window of 240-395ms after picture onset. Dashed lines above the x-axes indicate when decoding accuracy was reliable for the nouns (red) and adjectives (blue). Blue shading indicates the intervals during which the main effect of context was observed, that is, higher decodability of both categories when occurring in two-word contexts (phrase or list). Though not shown, there is an interaction effect 100–190 ms.
Fig 4.
Within-condition TGMs showing the temporal generalizability of noun (A) and adjective (B) representations from training time X to testing time Y in the three contexts. When nouns occurred in phrases, their representations generalized between earlier and later time-points in a way that was not observed for nouns in non-phrasal contexts or for adjectives in any context. This is evidenced by the off-diagonal instances of reliable decoding in the Noun in Phrase results (A, left). The right-most column shows subtractions between phrasal and non-phrasal contexts, with black boxing indicating significant differences.
Fig 5.
Across-condition TGMs showing the decodability and temporal generalizability of isolated word representations to phrase and list contexts.
Classifiers were trained on nouns and adjectives as they occurred in the isolation context and then tested when those same words occurred within phrases or lists. (A) Neural representations of nouns were sufficiently similar in isolation and in phrases such that decoding was reliable starting at 100ms and lasting till almost the end of the epoch. These representations also showed temporal generalizability starting at 100ms. Representations active at 100-200ms disappeared and then re-emerged about a hundred milliseconds later, while representations at 300-400ms stayed active in a more sustained fashion until the end of the epoch. (B) Adjective representations, in contrast, did not generalize from isolated contexts to phrasal contexts nearly as robustly. Mainly, shared representations across these two contexts were observed in a late time-window, close to articulation, at 500-600ms. This could reflect planning of the adjective articulation, which was the first word to be uttered in all three depicted contexts. In a similar late time-window, isolated adjective representations generalized to adjectives in the lists, though more weakly.