The role of syntax in semantic processing: A study of active and passive sentences

Background As you read this sentence, you are performing the complex task of composing the meaning of these words into a whole proposition. A multitude of work in psycholinguistics has shown that humans use both the semantics (the meaning of the words) and the syntax (their order) of a sentence to process its meaning. However, the rules by which the brain combines this information are largely unknown. High temporal resolution non-invasive neuroimaging techniques, such as magnetoencephalography (MEG) have great potential to answer this question, because they measure neural activity on the timescale of natural reading. Aim Using a machine learning approach, we contrast sentences with the same semantics and different syntax in order to shed light on the processing stages used by the brain when reading English sentences. We examine the robustness of these results to various analysis design choices to maximize sensitivity in future confirmatory analyses. Data We collected MEG data from 8 participants while they read activeand passive-voiced versions of the same propositions. Each participant read 16 propositions made up of 8 nouns and 4 verbs. The 32 sentences were repeated 15 times and we use an average of some or all of these repetitions for subsequent analysis. Methods We apply a Naive Bayes classifier at each point in the MEG activity timeseries to generate classification accuracy timeseries for each word in the sentence, cross-validating in a leave-one-proposition-out manner. We seek time windows of above chance accuracy in order to assert when certain words are accessed. To test the robustness of the observed effect, we vary different aspects of the data analysis and observe the effect on decoding accuracy. We also assessed which neural regions contribute most to decoding accuracy at different times. Results When reading passive-voiced sentences, participants access words relevant to the proposition multiple times throughout the sentence. This draws a sharp distinction with the processing of active-voiced sentences, in which participants seem to access relevant words only once, at word presentation. The result is highly robust to a variety of analysis decisions, and seems, to our surprise, to originate in occipital regions of the brain. Conclusions This result is consistent with prior work demonstrating reanalysis in syntactically complex sentences, in which participants correct an erroneous initial reading of the sentence. Our work elucidates the timing of reanalysis, which is late in the sentence and is consistent with theory on the mental storage of sentence components by role. Via reanalysis, participants can arrive at the same meaning through different syntactic paths. These results demonstrate the potential of machine learning approaches to reveal neural processing stages in an interpretable manner.