Write a PREreview

We present ME-Net (Morpho-Evidential Network), a compact dual-stream architecture for Indonesian hoax detection, together with a closed-form specification of the linguistic evidence it consumes. One stream is a convolutional n-gram encoder over tokens; the other is a parameter-free feature map that measures lexicon-verified affix rates, rumour versus institutional evidentials, register markers and surface statistics. A complementary gate computed from both views decides, per document and per hidden dimension, how much of the fused representation comes from each stream, and its average ¯g is an interpretable text-reliance score read directly off the forward pass. The contribution is the design and its analysis rather than a leaderboard result. We give exact definitions of all 13 features, including an affix predicate with allomorph restoration that avoids the systematic over-counting of naive prefix stripping, and we prove three properties: rate features are invariant under document duplication while the two count features are not, which separates length-dependent from length-invariant evidence; naive concatenation is exactly the case Wg = 0 of the gate, so gating can only enlarge the hypothesis class, with the gated multimodal unit recovered under a further constraint; and the gate rescales the two streams’ Jacobians by g and 1−g, so it performs per-dimension credit assignment during training. The symbolic stream and the gate together add 0.10M parameters (2.4%) over a text-only CNN and are Θ(N)atinference. We also report, briefly and without overstating them, the preliminary findings from our own implementation, because they bear directly on the design: on a saturated in-domain benchmark the gate converges to a nearly constant mixing weight, i.e. the collapse predicted by the expressiveness result, and the accuracy of the gated model is then indistinguishable from plain concatenation. The measurable benefits we observed lie elsewhere: markedly better data efficiency in the low-resource regime and comparatively stable behaviour under corpus shift. A full empirical study is left to future work; this paper fixes the architecture, the feature map and the analysis they rest on.