Lexis-E: Linguistically-Structured Hierarchical Text Compression and the Structural Compressibility of English

We present Lexis-E, a linguistically-structured hierarchical text compressor for English that achieves 2.7523 bits-per-byte (bpb) on FineWeb and 2.7555 bpb on Moby Dick at 100,000 characters, using no learned weights and no training corpus.

Lexis-E operates through a 12-stage pipeline that progressively strips linguistic redundancy at every level of English structure -- morphological, syntactic, discourse, phonetic, and probabilistic -- encoding text as sequences of linguistically-motivated symbolic deltas fed into an online context-mixing model and arithmetic coder.

Against classical compressors on the same corpora, Lexis-E outperforms gzip (3.3948 bpb), zstd (3.1828 bpb), and xz (3.1079 bpb) on the character stream, purely through explicit linguistic structure. Key technical contributions include mixed-radix phonetic decomposition, which reduces character-level delta magnitude by 2.42×; factoriadic delta encoding; anchor-based symbol slot splicing for zero-overhead discourse token handling; and a configurable compact metadata envelope that reduces full-payload overhead by 47–52% versus the baseline implementation.

Using these results, we quantify the contribution of linguistic structure to English compressibility: the total gap between unencoded UTF-8 (8.00 bpb) and GPT-2 1.5B (≈1.30 bpb) is 6.70 bpb, of which Lexis-E closes 5.25 bpb -- approximately 78% -- without any statistical learning.

This suggests that a large fraction of what neural language models learn about English is recoverable from explicit linguistic priors alone, and motivates further investigation into the boundary between structural and statistical compressibility.