{"id":"18cfe662-8832-4667-ba72-21368b813de7","arxiv_id":"2412.15250","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Removing vowels before LZW compression yields high compression ratios, but the resulting text cannot be restored without a large transformer, making the claimed state-of-the-art comparison unfair.","lead":"The paper removes all vowels from English text, compresses the vowel-less text with LZW, and uses a transformer model to restore the vowels. The authors claim state-of-the-art compression ratios, but the comparison ignores that the method is lossy and that the decompression model is not included in the compressed size.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA compression ratios omit the decompression model and treat a lossy, vowel-stripped representation as the original text, so the central claim is not supported by the reported metric.","rationale":"The reader's weakest assumption identifies the same load-bearing defect: the compression ratio in Eq. 1 and Section V.C is defined over LZW(vowel-removed text), not over the original text, and the 63M-parameter transformer needed for reconstruction is excluded. This is a correctness risk, not merely a stylistic or presentation issue, because it directly invalidates the state-of-the-art comparison to lossless compressors in Table II. The proposed concrete test would settle the concern by adding the model cost and checking exact reconstructability. The paper provides no code, data, or formal verification, and the reported BLEU scores confirm that reconstruction is imperfect, so an UNCHANGED REJECT verdict is appropriate.","tokens_in":9626,"tokens_out":2731,"duration_ms":26422,"concrete_test":"Recompute the compression ratio on the actual 100K-sentence test corpus using total storage bits = |LZW(vowel-removed corpus)| + |decompression model| (e.g., 63.23M parameters × 16 bits). Compare corrected_ratio = original_text_bits / total_storage_bits against GZIP, LZMA, and the GPT-based baseline on the original text. If corrected_ratio falls below those baselines, the reported SOTA is an artifact of excluding the model. Separately, report the exact-match rate after vowel restoration; if it is below 100%, the method is lossy and cannot be benchmarked against lossless compressors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V.C defines compression ratio as Original Length / Compressed Length, and Table II reports ratios computed on the LZW-compressed, vowel-removed text (Section IV.A, Eq. 1). The vowel-removed string cannot be turned back into the original source without RejuvenateFormer or T5-Small, a 63M-parameter transformer whose storage is never counted in the compressed size. Thus the reported ratios of 12.57, 13.38, and 11.42 are not compression ratios for the original text; they describe a lossy, model-dependent representation. Comparing those numbers to GPT-based lossless compression and to traditional lossless compressors is therefore apples-to-oranges. A fair comparison must either include the decompressor model's cost (e.g., its parameter bits) or evaluate rate-distortion against the achieved BLEU/BERTScore; the paper does neither. Since the central SOTA claim rests entirely on this denominator choice, the claim is unsupported as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a text-compression pipeline in which vowels are first removed from English text, the vowel-less text is compressed with the Lempel-Ziv-Welch (LZW) algorithm, and a transformer-based model (RejuvenateFormer or a pre-trained T5-Small) is used to restore the missing vowels. The central claim is that this pipeline achieves compression ratios of 12.57, 13.38, and 11.42 on BookCorpus, WMT14 EN-DE, and WMT14 EN-FR, respectively, and that these are state-of-the-art compared with both deep-learning and traditional compressors. The paper also reports restoration quality in terms of BLEU and BERTScore, and includes an ablation study on training-set size. The key evaluation issue is that the reported compression ratio is computed as the original text length divided by the LZW-compressed length of the vowel-removed text, without accounting for the fact that the compressed representation alone cannot reconstruct the original text.","tokens_in":9810,"tokens_out":4599,"duration_ms":42477,"significance":"If the reported ratios were genuine end-to-end compression ratios, the result would be a meaningful advance in neural text compression. The paper does provide a concrete experimental setup, compares several standard compressors after vowel removal, and reports restoration-quality metrics, and the corpus-size ablation in Table III is a reasonable check. However, the central claim is not supported by the metric as defined: the denominator excludes the decompression model, and the method is lossy because vowel removal is irreversible and restoration is approximate. These two omissions make the reported state-of-the-art compression ratios an artifact of the evaluation definition rather than a property of a complete compression system.","major_comments":[{"comment":"The compression ratio is defined as Original Length / Compressed Length, but the compressed length is the LZW representation of the vowel-removed text. This is not the compressed size of the original text: the original cannot be recovered from the LZW output alone, because vowel removal is irreversible and the transformer model is required to reconstruct the missing characters. The model size is never counted in the compressed representation. A concrete estimate shows the impact: for BookCorpus, the reported compressed size is roughly 6.7M / 12.57 bytes, or about 4.26 Mb, while RejuvenateFormer has 63.23M parameters, which is about 2.02 Gb at 32 bits per parameter, so including the model makes the effective end-to-end compression ratio less than 1. A fair metric must either include the decompressor's cost or evaluate rate-distortion against the achieved reconstruction quality.","section":"Section IV.A, Eq. (1); Section V.C"},{"comment":"The comparison to lossless compressors is not apples-to-apples. The proposed pipeline is lossy: vowels are removed, and the transformer restores them only approximately, as shown by BLEU scores of 27.31, 25.78, and 50.45 and F1 scores around 0.89 to 0.95 in Table I. In contrast, LZMA, GZIP, ZLIB, arithmetic coding, and the GPT-based and TRACE baselines preserve the original text exactly. Reporting a higher compression ratio for a lossy scheme without a rate-distortion comparison, and calling it a state-of-the-art compression ratio, is misleading. The manuscript should either compare against lossy methods on the same terms or clearly present the result as a lossy rate-distortion trade-off.","section":"Section V.D.1, Table II"},{"comment":"The manuscript repeatedly describes the method as lossless compression, specifically in the Abstract ('a lossless compression method') and in Section IV.B.1 where LZW is described as a lossless algorithm. While LZW itself is lossless, the complete pipeline is lossy because vowel removal is not invertible and the transformer restoration is not exact. This terminology obscures the central issue: the compressed representation stores a vowel-less text that does not determine the original text. The paper should consistently refer to the approach as lossy text compression with learned restoration, and should state the reconstruction error explicitly as part of the compression claim.","section":"Abstract; Section IV.B.1"}],"minor_comments":[{"comment":"The abstract writes 'RejuvenateForme' once, while the rest of the paper uses 'RejuvenateFormer'; please fix the typo.","section":"Abstract"},{"comment":"The column header 'GLIB' appears to be a typo for 'ZLIB'.","section":"Table II"},{"comment":"The BBFNMT row reports BLEU scores from the original machine-translation task, while the other rows report vowel-restoration BLEU scores; these are not directly comparable and the table should state this limitation.","section":"Table I"},{"comment":"The text attributes T5-Small's advantage to 'training on a larger corpus', but T5-Small is pre-trained on C4 rather than on the datasets used here; the explanation is unclear and should be revised.","section":"Section V.D.2"},{"comment":"The notation in Eq. (2) is ambiguous: the decoder input $D_{t-1}^{out}$ and the weight matrices $W_E, W_D$ are not defined in the text. Please define all symbols used in the equation.","section":"Section IV.A, Eq. (2)"}],"recommendation":"reject","confidential_remarks":"The central compression-ratio claim is an artifact of an unfair metric that omits the decompression model and ignores the lossy nature of the reconstruction. The authors could potentially reframe the work as a lossy neural compression method with a rate-distortion evaluation, but as submitted the main claim is not supported by the reported experiments. The comparison to lossless compressors and to prior neural methods is likely to mislead readers, so the paper is not suitable for publication in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline result here is the compression ratio, and it does not hold up. The paper strips vowels, runs LZW, and then reports original length divided by the LZW-compressed vowel-less string. That denominator cannot reconstruct the original without the 63M-parameter transformer, whose storage is never counted. So the 12.57–13.38× numbers are not compression ratios for the original text; they are ratios for a deliberately degraded representation. Comparing them to lossless compressors and to GPT-based lossless schemes is apples-to-oranges.\n\nWhat is genuinely new is small but real: the specific pipeline of vowel removal + LZW + transformer restoration, applied to three corpora, with restoration quality measured by BLEU and BERTScore. The paper is honest that the restoration is lossy and reports those scores rather than hiding them. The ablation on corpus size is useful, and the comparison with T5-Small is a reasonable sanity check. The authors also do not oversell RejuvenateFormer; they acknowledge T5-Small performs better.\n\nThe soft spots beyond the metric are noticeable but secondary. There is no code or data release, no error bars, and the writing is rough with typos and a few table inconsistencies. The ablation is only on EN-DE. The novelty is modest: vowel removal as a compression pre-processing trick appears in earlier lossy work, and transformer restoration is established. The contribution is a variation, not a new paradigm.\n\nI would not publish this as a state-of-the-art compression result. The central claim rests on an unfair comparison, and a fair rate-distortion evaluation—including the model's cost, or measuring bits-per-character against reconstruction quality—would, I suspect, shrink the advantage considerably. That said, the idea is not worthless. Reframed as a lossy compression-then-restoration system with proper rate-distortion reporting, it might be a small workshop contribution.\n\nMy recommendation: desk reject as is, or send back for a major revision that fixes the metric. It does not deserve a full referee cycle with the current framing.","headline":"The claimed SOTA compression ratios fail because they compare original text length against LZW-compressed vowel-stripped text while omitting the 63M-parameter restoration model, so the central result is unsupported; the restoration experiments are honest but the paper is not ready for publication as is.","tokens_in":10291,"tokens_out":2933,"would_cite":false,"duration_ms":26232,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that deleting all vowels from English text and compressing with LZW reaches compression ratios of 11.42 to 13.38 on three corpora, with transformers able to restore the missing vowels.","keywords":["text compression","vowel removal","Lempel-Ziv-Welch","transformer","text restoration","BLEU","BERTScore"],"falsifier":"Take a fixed English corpus, encode it as the compressed vowel-free text plus the weights of the restoration model, divide the total bytes by the original size, and compare with a standard lossless compressor such as GZIP on the same corpus; if the end-to-end number is not smaller, the central compression-ratio claim is falsified.","tokens_in":9425,"feed_emoji":"🗜️","tokens_out":6940,"duration_ms":57596,"temperature":0.7,"pith_summary":"This paper is trying to establish that a simple preprocessing step, deleting every vowel from English text, changes the compression problem enough that a standard LZW compressor reaches compression ratios around 12 to 13 times on BookCorpus, WMT14 EN-DE, and WMT14 EN-FR. A transformer-based model called RejuvenateFormer, or the pretrained T5-Small, then reconstructs the missing vowels with moderate to high BLEU and BERTScore. If true, the practical consequence is that English text can be stored or transmitted at roughly eight percent of its original byte size on the paper's metric, at the cost of needing a neural restoration stage. The authors also claim that the preprocessing is compatible with any transformer encoder-decoder and that larger training corpora improve restoration quality.","feed_headline":"Stripping vowels boosts text compression to 13x","feed_subtitle":"A transformer fills the vowels back in, reaching BLEU 50 on BookCorpus and 27 on WMT14 EN-DE.","key_machinery":"The load-bearing object is a three-stage pipeline written as Equation (1): $S_{RCI} = LZW_D(LZW_C(VR(S)))$, where $VR(\\cdot)$ removes the ten English vowel characters, $LZW_C$ compresses the vowel-free string losslessly and defines the reported compression ratio, and $LZW_D$ returns the vowel-free text. A transformer decoder then generates the missing vowels autoregressively using multi-head self-attention over the reconstructed vowel-free sequence. This setup separates the lossless part, handled by LZW, from the lossy part, vowel deletion plus neural restoration, which is what makes the high compression ratios possible.","core_discovery":"The central claim is that vowel removal is an effective lossy front-end for English text compression. The paper reports compression ratios of 12.57 on BookCorpus, 13.38 on EN-DE, and 11.42 on EN-FR, computed as the original text length divided by the LZW-compressed length of the vowel-removed text, and presents these as beating a GPT-based compression approach, TRACE, and general-purpose LZMA, GZIP, ZLIB, and arithmetic-coding baselines. The discarded vowels are then treated as a sequence-generation task: RejuvenateFormer, a six-layer encoder-decoder transformer with a 512-dimensional hidden state, is trained to map vowel-free sequences back to complete English sentences, achieving BLEU scores of 50.45 on BookCorpus, 27.31 on EN-DE, and 25.78 on EN-FR. The paper's own results also show that the pretrained T5-Small restores vowels more accurately than RejuvenateFormer on all three corpora, which the authors take as evidence that the preprocessing is model-agnostic.","pith_inferences":["A fair end-to-end comparison would count the restoration model's parameters, and arguably its inference compute, as part of the compressed representation; under that accounting the reported 11.4 to 13.4 times ratios would shrink, and the comparison against truly lossless codecs would be less favorable.","The same pattern, delete a predictable character class, compress, then regenerate, could be ported to diacritics in Arabic or Hebrew, to punctuation, or to other high-redundancy signals, with the restoration difficulty set by how much the deleted class constrains the original text.","BLEU scores likely overstate recoverability in this setting because fluent, plausible text can score well without matching the original sentence exactly; exact-match rate or character error rate would be a stricter test of restoration fidelity."],"forward_implications":["On the paper's own metric, English text can be represented at roughly eight percent of its original size by storing only the LZW-compressed vowel-free form and relying on a restoration model at use time.","Because T5-Small outperforms the from-scratch model, the method does not depend on RejuvenateFormer's specific architecture; any adequate sequence-to-sequence transformer can act as the vowel restorer.","Restoration quality improves with corpus size: on EN-DE, BLEU rises from 23.07 with 30K training pairs to 27.31 with 100K pairs, so the approach is expected to benefit from larger-scale training.","The method is language-specific as presented: the removed character set is exactly the five English vowels in upper and lower case, so applying it to other languages would require redefining that set."],"supporting_citations":[{"why":"Supplies the LZW algorithm used for the lossless compression half of the pipeline.","marker":"[3]"},{"why":"Provides the TRACE transformer-based lossless compressor that the paper compares against.","marker":"[11]"},{"why":"Provides the GPT-based compression approach that the paper claims to outperform.","marker":"[12]"},{"why":"Defines the multi-head self-attention and positional encoding on which RejuvenateFormer rests.","marker":"[23]"},{"why":"Supplies the WMT14 EN-DE and EN-FR corpora used for training and evaluation.","marker":"[24]"},{"why":"Supplies the BookCorpus dataset used for training and evaluation.","marker":"[25]"},{"why":"Provides the T5-Small pretrained model that serves as a baseline and the best-performing vowel restorer.","marker":"[26]"},{"why":"Defines BERTScore, one of the two evaluation metrics used to measure restoration quality.","marker":"[27]"},{"why":"Defines BLEU, the other evaluation metric used to measure restoration quality.","marker":"[28]"},{"why":"Supplies arithmetic coding as a general-purpose lossless compressor baseline.","marker":"[31]"}],"fun_headline_variants":["Vowel-free text compresses 13x, AI restores it","Drop vowels for 13x compression, model fills them in","13x compression: strip vowels, let transformer regenerate","Compress text 13x by removing vowels, model rebuilds them","Vowel removal boosts compression 13x, transformer recovers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported compression ratio divides the original text length by the compressed vowel-free text length, without counting the size of the restoration model or the fact that the vowel-free text alone cannot be turned back into the original text; on a fair end-to-end metric this advantage would shrink.","fun_headline_variants_meta":{"raw":{"variants":["Vowel-free text compresses 13x, AI restores it","Drop vowels for 13x compression, model fills them in","13x compression: strip vowels, let transformer regenerate","Compress text 13x by removing vowels, model rebuilds them","Vowel removal boosts compression 13x, transformer recovers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000322,"raw_usage":{"total_tokens":1843,"prompt_tokens":1011,"completion_tokens":832,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":744}},"tokens_in":627,"tokens_out":832,"duration_ms":6121,"temperature":1.0,"reasoning_tokens":744,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:21:44.423290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed English corpus, encode it as the compressed vowel-free text plus the weights of the restoration model, divide the total bytes by the original size, and compare with a standard lossless compressor such as GZIP on the same corpus; if the end-to-end number is not smaller, the central compression-ratio claim is falsified.","supporting_citations":[{"cited_title":"Lzw data compression,","cited_arxiv_id":null,"evidence_quote":"Supplies the LZW algorithm used for the lossless compression half of the pipeline."},{"cited_title":"Trace: A fast transformer- based general-purpose lossless compressor,","cited_arxiv_id":null,"evidence_quote":"Provides the TRACE transformer-based lossless compressor that the paper compares against."},{"cited_title":"Findings of the 2014 workshop on statistical machine translation,","cited_arxiv_id":null,"evidence_quote":"Supplies the WMT14 EN-DE and EN-FR corpora used for training and evaluation."},{"cited_title":"Aligning books and movies: Towards story-like visual explanations by watching movies and reading books,","cited_arxiv_id":null,"evidence_quote":"Supplies the BookCorpus dataset used for training and evaluation."},{"cited_title":"Bertscore: Evaluating text generation with bert,","cited_arxiv_id":null,"evidence_quote":"Defines BERTScore, one of the two evaluation metrics used to measure restoration quality."},{"cited_title":"A call for clarity in reporting BLEU scores,","cited_arxiv_id":null,"evidence_quote":"Defines BLEU, the other evaluation metric used to measure restoration quality."}],"review_version":1}