Anling Xiang

Jingyang Du, Yang Shen, Anling Xiang

Burrows-Wheeler-transform-based compressors rely on local context regularity, but structured text also contains dates, counters, identifiers, coordinates, and other digit runs whose values vary differently from their surrounding tokens. STC is a practical BWT-family compressor that separates this source of variation before the component BWT stage. It replaces digit runs in the main stream with an unambiguous placeholder and stores the removed digits in length- and context-conditioned side streams. The side streams use stable bucket ordering and compact digit packing, so the decoder can reconstruct the original run order from the normalized main stream without storing a separate permutation. The resulting components are encoded by a fixed internal BWT/M03-style component coder. On enwik9, STC produces a 157,388,188-byte archive with a 183,174-byte decoder source package, giving a local LTCB-style total of 157,571,362 bytes. A full-enwik9 same-coder ablation shows that the digit-context decomposition reduces the archive by 2,629,561 bytes relative to the no-split control. The result is locally verified by full decode and SHA-256 matching; official benchmark status requires independent maintainer-side verification.

Anling Xiang, Yuwen Yang, Yang Shen

Scholarly text classification supports literature organization, subject indexing, and research intelligence, but Chinese scholarly corpora often contain imbalanced and semantically adjacent disciplinary labels. We propose AutoTail-BSFGM, a class-balance-aware fine-tuning method that combines an automatically gated tail-prior adjustment, a weak Balanced Softmax auxiliary loss, and Fast Gradient Method adversarial regularization. The method changes only the training objective and procedure; inference uses the same single base-size encoder and linear classifier as the corresponding label-smoothed baseline. We evaluate the method on two CSL-based tasks: an abstract-to-discipline task with 67 labels and a title-to-category task with 13 categories. On the primary abstract task, AutoTail-BSFGM improves validation and lockbox accuracy under both Chinese RoBERTa-WWM and MacBERT-base. With MacBERT-base, validation accuracy increases by 0.83 percentage points and lockbox accuracy by 0.49 points, with a pooled paired McNemar signal on validation (p = 0.023). On the title task, the method improves validation accuracy by 0.70 points and validation balanced accuracy by 2.64 points; lockbox accuracy is approximately neutral while lockbox balanced accuracy improves by 1.22 points. The results support a bounded contribution: AutoTail-BSFGM improves class-balance-sensitive behavior and yields consistent gains for abstract-based scholarly classification, without uniformly improving every metric on every split.