Intelligent Translation for Less Commonly Taught Languages: Symmetry-Aware Bidirectional Decoding with Linguistic Guidance and XLNet Enhancement

Autoregressive Transformer translation relies on a target prefix and does not directly condition each prediction on an ungenerated target suffix. This study evaluates how complementary decoding directions, linguistic guidance, and pretrained source representations are integrated in a family of bidirectional translation configurations. BiDNMT uses an auxiliary right-to-left decoder; TeF-BiDNMT examines scheduled target replacement; KDS-BiDNMT combines keyword and dependency information; and XLK-BiDNMT links XLNet representations to serial–parallel cross-attention. These information-flow choices are of potential interest to translation involving less commonly taught languages, where contextual evidence may be uneven, but their effectiveness is evaluated here only on WMT17 Chinese–English and WMT14 English–German benchmarks. At beam size 5, BiDNMT records 22.42/29.96 BLEU and TeF-BiDNMT 26.02/25.65 on Chinese–English/English–German, respectively. KDS-BiDNMT is reported at 23.39/31.12 BLEU in its main comparison, and XLK-BiDNMT at 27.87, 23.61, and 27.12 across newstest2017, newstest2018, and cwmt2018. The length-stratified comparison shows both positive and negative differences, with the largest positive margin of 6.40 BLEU points in (70,80]. The evidence therefore characterizes condition-dependent architectural behavior; it does not establish improved performance on less commonly taught languages or uniform superiority on the evaluated language pairs.