Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Poster Slams
Language system contributes to ‘gist’ extraction during code comprehension
Poster A42 in Poster Session A, Thursday, October 6, 10:15 am - 12:00 pm EDT, Millennium Hall
Also presenting in Poster Slam A, Thursday, October 6, 10:00 - 10:15 am EDT, Regency Ballroom
Yun-Fei Liu1, Marina Bedny1; 1Johns Hopkins University
Programming languages, such as Python and C++, are a recent cultural invention that “recycles” cortical systems that evolved for other purposes. Natural and computer languages share some features: They use common symbols (letters and words), both have hierarchical structure, and both are recursive (Fedorenko et al., 2019, TiCS). Does computer programming reuse cortical networks that evolved for language processing? Two recent studies, including one from our lab, suggest that programming does not recycle fronto-temporal language systems and instead makes use of fronto-parietal executive/logical reasoning systems (Liu et al., 2020, eLife; Ivanova et al., 2020, eLife). However, there is strong co-lateralization between the code and the language systems across individuals and some, albeit small, overlap between language- and code-comprehension (Liu et al., 2020, eLife). We conducted follow-up time-course and MVPA analyses to ask whether the language network is involved in the initial stages of extracting meaning from code, whereas the fronto-parietal logic network computes the algorithms. During an MRI scan, expert programmers (mean years of experience=5.7, n=15) read Python functions containing either FOR or an IF control structures, followed by an input, then an output. Participants judged whether the output was correct. Participants performed a memory control task with the same FOR or IF functions but with the words presented in scrambled order, rendering the function meaningless. The same participants also performed a separate language/logic localizer scan where they judged whether two sentences have the same meaning (language) or whether two logical statements are consistent (logic. e.g., if both X and Y then not Z, if Z then not either X or Y?) (Monti et al., 2009, PNAS). As previously reported, in univariate analysis, the code-responsive network identified by real > scrambled code overlapped more with logic than language. In a time-course analysis, the fronto-parietal logic network showed a robust peak of activation around 15 seconds after the onset of code stimuli, consistent with its involvement in the algorithmic processing of code. By contrast, lateral temporal language areas showed a small peak of activation earlier, 5 seconds after code function onset. Surprisingly, multivariate decoding of FOR vs. IF functions was just as robust in lateral temporal language areas (72.3%) as fronto-parietal logic areas (PFC 64.7%, IPS 67.4%). Decoding in language areas was evident in both ROI and whole-brain analyses, showing overlap between language localizer and FOR vs IF function decoding. Could language areas distinguish between FOR and IF functions based on the presence of different words (e.g., presence of the keywords “for” and “if”)? Contrary to this idea, we could not decode scrambled FOR from scrambled IF functions. This suggests that decoding in language areas is not driven by word-level differences. We hypothesize that the language system plays a role in the initial stage of code comprehension during which “gist” information is extracted but the algorithm underlying the code is not yet parsed. Later, the information extracted by the language system is transmitted to the logical reasoning system, where the algorithm is understood.
Topic Areas: Meaning: Combinatorial Semantics, Reading