ReSTA Net: Recursive Summarizer–Transcoder–Analyzer Network for Tree-Structured Text Processing

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We present ReSTA Net (Recursive Summarizer–Transcoder–Analyzer Network), a neural architecture that encodes a sentence into a single fixed-dimensional vector along an induced binary syntax tree and decodes the sentence back from that vector alone. A Syntax Tree Parser induces the tree by multi-scale sliding-window voting; a Summarizer fuses child vectors bottom-up; a deep Transcoder maps the semantic summary of the input sentence to that of the output sentence; and an Analyzer recursively splits vectors top-down, stopping when a decoded vector is close enough to a vocabulary embedding. Training follows the target tree by teacher forcing with a mean-squared-error objective, and two training strategies are described. Experimental results will be reported in a follow-up paper.