{"id":"22add3f7-f2a6-417b-a309-801b3a0ca474","arxiv_id":"1909.05364","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"TransSent generates a tail discourse from a head discourse and a discourse marker by treating the marker as a translation in embedding space, with new datasets and improved scores over baselines.","lead":"TransSent generates the second half of a structured sentence from the first half plus a linking word such as 'but' or 'because'. It treats the linking word as a mathematical translation between sentence meanings, and reports better automatic and human scores than several text generation baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Eq. 10 objective never trains the decoder on the translated representation: Lrec only reconstructs from true encodings zt, while Ldis/Lratio act in relation space, so the generation path zt*→s_t* receives no gradient and rests on an untested closeness assumption.","rationale":"The reader's weakest assumption concerns the shared BERT encoder/discriminator potentially inflating ACC scores. That is a valid metric-fairness concern, but it is secondary. The more fundamental issue is internal to the stated method: the training objective in Eq. 10 does not train the component of the model that actually generates text from the translated representation. The decoder is trained exclusively through Lrec on the true encodings zt and zh, while the relation path's output zt* is decoded at inference time by that same decoder, which has never been optimized to decode such vectors. The inverse feed-forward network that maps zt*_r back to discourse space also receives no gradient from any term in Eq. 10. Thus the central mechanism of TransSent is unsupervised as written. This is not a matter of the evaluation metric being self-serving; even with a perfect external metric, the stated objective cannot explain how the model learns to produce grammatical, relation-correct tails. The empirical tables could still be accurate if the authors' implementation secretly included an additional generation loss, but no such loss appears in the paper, Algorithm 1, or the released artifacts, and no code is available to confirm. Because the central claim depends on an untested and unstated assumption about relation-space closeness transferring to decodability, the paper as written does not support its headline claim. A direct test—adding a supervised generation loss on zt* and comparing results—would settle whether the published objective is merely incomplete or whether the relation-space losses alone suffice. Until that test is run, the correctness of the method is unestablished, so the appropriate verdict is REJECT rather than CONDITIONAL.","tokens_in":9848,"tokens_out":6223,"duration_ms":75968,"concrete_test":"Compute the gradient of Eq. 10 with respect to the decoder parameters Gdec along the path through zt*: because Lrec uses zt and Ldis/Lratio use zt*_r before the inverse mapping, this gradient is identically zero. Then implement Algorithm 1 exactly and verify whether the decoder actually receives any training signal for decoding translated representations; alternatively, add Lgen = -log p(st|zt*) to Eq. 10 and retrain on Book-dm. If ACC, PPL, and human ratings do not improve over Table 3, the current relation-space objective is sufficient; if they improve substantially, the published objective omits the term actually responsible for generation quality.","verdict_should_be":"REJECT","load_bearing_attack":"In Section 4.2, the relation path produces zt*_r via Eq. 7, Ldis is computed by Eq. 8 and Lratio by Eq. 9, both in relation space, and then a feed-forward network maps zt*_r back into discourse space as zt*, from which s_t* = Gdec(zt*) is decoded. However, the total objective in Eq. 10 is L = Lrec + λd Ldis + λr Lratio, where Lrec (Eq. 4) is only -1/2(log p(sh|zh) + log p(st|zt)). No term evaluates log p(st|zt*) or otherwise supervises the decoder or the inverse mapping on translated representations. As a result, gradients from Ldis and Lratio update only the relation-translation parameters; the decoder and the inverse feed-forward network are trained only on genuine BERT encodings zt (and zh). At inference, the decoder is asked to decode zt* values drawn from a distribution it has never been trained on. The paper implicitly assumes that closeness in relation space (small Ldis) plus an untrained inverse map implies closeness in discourse space, so that the decoder's language-modeling skill transfers to zt*. That assumption is neither stated nor tested, and it is the load-bearing condition for the central claim that TransSent generates coherent, relation-correct tails. Without it, the high ACC numbers could reflect relation-space geometry rather than the quality of generated text, and the human ratings (~3/5) are too coarse to rescue the mechanism. This is an internal gap in the training objective, not a disagreement with external consensus.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a new task, discourse transfer, in which a model receives a head discourse and an explicit discourse marker (and, but, because, if, when) and must generate the tail discourse of a structured sentence. The authors propose TransSent, an encoder-decoder model with a frozen BERT encoder, a unidirectional LSTM decoder, and a relation translation network inspired by TransR. The relation network projects head and tail discourse representations into a separate relation space, performs a translation-like operation using the discourse-marker embedding, and maps the result back to discourse space for decoding. The training loss in Eq. (10) combines a reconstruction loss on genuine head/tail encodings with two relation-space losses. The paper reports experiments on three datasets (Yelp-dm, Wiki-dm, Book-dm) for discourse transfer, free text generation, and dialogue generation, and claims that TransSent produces structured sentences with high quality, based on automatic metrics (ACC, NLL, PPL) and human ratings.","tokens_in":10208,"tokens_out":4816,"duration_ms":55779,"significance":"If the proposed mechanism worked as described, explicit discourse-marker-conditioned generation would be a useful step toward controllable and structured text generation, and the analogy to knowledge-graph translation would be an interesting architectural idea. The three constructed datasets and the task formulation are potentially reusable resources. However, the central training objective does not supervise the actual generation path (decoding from translated representations), and the automatic evaluation metric uses the same fine-tuned BERT as the model encoder. These are load-bearing issues in the current manuscript: the strong quantitative claims are not yet supported by the described method. The significance would be substantially improved by adding a loss that trains the decoder on translated representations, reporting an independent automatic metric, and providing ablations that separate the contribution of the relation translation from the benefit of the pretrained BERT encoder.","major_comments":[{"comment":"The training objective never supervises the decoder on the translated representation. In Eq. (4), Lrec only reconstructs sh and st from their genuine encodings zh and zt; Ldis in Eq. (8) and Lratio in Eq. (9) are computed entirely in relation space (between zt*r and ztr, and between zt*r and zhr). The feed-forward inverse mapping that produces zt* and the decoder Gdec that produces st* receive no gradient from these terms. Therefore, at inference the decoder is asked to decode vectors zt* that are drawn from a distribution on which it was never trained. This is the load-bearing path for the paper's central claim that TransSent generates coherent, relation-correct tails. A loss such as log p(st | zt*) or an explicit distributional alignment between zt* and zt is needed; alternatively, the authors should provide an experiment or analysis showing that the untrained decoding path nevertheless works.","section":"§4.2, Eq. (10)"},{"comment":"The automatic ACC metric uses the same fine-tuned BERT that serves as the frozen encoder in §4.1. Because that BERT was fine-tuned on the DMP task and then used as the discriminator, ACC is not an independent measure of discourse correctness; it measures agreement with the model's own representation learner and can therefore overstate the relation accuracy of TransSent outputs. The human evaluation is too small to compensate: five annotators rate 200 examples per model/dataset, with no inter-annotator agreement and no confidence intervals, and the average ratings (3.0-3.4 out of 5.0) are modest. The authors should report ACC with a separately trained classifier and provide human-evaluation statistics, e.g., confidence intervals or agreement measures.","section":"§6.3, Table 3"},{"comment":"All baselines use random-initialized RNN/LSTM encoder-decoder architectures, while TransSent uses a frozen BERT encoder. The large gains in ACC, PPL, and NLL may therefore be attributable to the pretrained encoder rather than to the proposed relation translation network. No ablation is reported that isolates the contribution of the relation translation, such as a version of TransSent without the relation path or with a simple additive translation in discourse space. An ablation is needed to support the claim that translation in relation space is the source of improvement.","section":"§6.1, §6.4"}],"minor_comments":[{"comment":"The notation p(st) in Eq. (3) is ambiguous: the decoder should condition on the encoded representation, so p(st | zh) or similar would be clearer. Also, Eq. (3) appears to define the decoder for zh only, while the same decoder is used for zt; this should be stated explicitly.","section":"§4.1, Eq. (3)"},{"comment":"There are several typos: 'nolinear' should be 'nonlinear' in §1 and §4; 'comparision' should be 'comparison' in §6.3; 'CA VE' should be 'CVAE' in §2 and §6.1.","section":"Throughout"},{"comment":"The paper states that beam search is not used, but it does not specify the decoding procedure (e.g., maximum-likelihood decoding or sampling). This detail is needed for reproducibility.","section":"§6.2"},{"comment":"The tables report mean scores without error bars or significance tests. Given the modest differences on some cells, statistical significance should be reported.","section":"Table 3 and Table 4"},{"comment":"In Algorithm 1, line 8 says 'Calculate Ldis and Lratio' and line 9 'Calculate L', but Eq. (10) is the total objective. It would be clearer to separate the network parameters that are updated by each loss term, since the decoder and inverse mapping are not updated by the relation-space losses.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The main concern is a concrete gap in the training objective: the decoder is never trained on the translated representation that is actually used at inference. This is fixable by adding a reconstruction or alignment loss on zt*, but as written the central mechanism is unsupported. If the authors cannot demonstrate that the untrained generation path works, the central claim would not hold. The shared-BERT evaluation and the lack of ablations also need to be addressed. Because the issues are internal and potentially fixable within a revision, I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nThis one is worth a look, but with a big caveat. The paper introduces discourse transfer — generating a tail clause from a head clause plus a discourse marker — and builds three datasets (Yelp-dm, Wiki-dm, Book-dm) from explicit discourse markers. That is a clean, reusable contribution, and the idea of borrowing TransR-style translation from knowledge-graph embeddings to relate discourse vectors is appealing.\n\nThe problem is in the training objective. Equation 10 is L = Lrec + λd Ldis + λr Lratio. Lrec reconstructs the original head and tail from their true BERT encodings. Ldis and Lratio operate in relation space on translated vectors. Nothing compares the translated vector z*_t (after the inverse feed-forward mapping back to discourse space) with anything, and no term evaluates the decoded text from z*_t. So the decoder and the inverse mapping never receive a gradient from the generation path that is actually used at test time. The paper implicitly assumes that small relation-space distance guarantees closeness after the untrained inverse map, but that is only an assumption. The decoder is asked to decode vectors from a distribution it has never seen. This is the load-bearing condition for the central claim, so it cannot be left unstated.\n\nThere are secondary issues. The ACC metric uses the same fine-tuned BERT that is the encoder; that shared representation space can inflate scores. Human ratings use only five annotators over 200 samples per model, with no significance testing. No code is released, so the discrepancy between the described training and the reported quality cannot be checked.\n\nI still think the task and datasets deserve attention, and the paper is clearly written. It should go to peer review, because the discourse-transfer task is genuinely new and could be a starting point for others. But an editor should send it to knowledgeable reviewers and expect the authors to either train the translated path or drop the claim that the relation translation is what makes the model work.\n\nMy recommendation: engage with it as a resource, not as a recipe. If you work on discourse-aware generation, cite the datasets once they are available, and treat the model description with suspicion. Reading-group maybe — it is a good example of a paper where the reported numbers likely outrun the mechanism.\n\nBest.","headline":"The paper contributes a genuinely new task and three datasets for discourse transfer, but the training objective never supervises the decoded translated tail, so the central generation mechanism is unverified and the reported numbers likely outrun the model description.","tokens_in":10725,"tokens_out":4179,"would_cite":false,"duration_ms":43877,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that generating a structured sentence—two clauses joined by a discourse marker—becomes easier when the model separates meaning from structure.","keywords":["discourse transfer","discourse markers","structured sentence generation","translation-based relation modeling","BERT","variational autoencoder","dialogue generation"],"falsifier":"Train an independent discourse-marker classifier on the same DMP data but with a different architecture and no exposure to TransSent training, then run TransSent's generated tails through it. If its accuracy stays near the reported 50–80% on open-domain data, the metric is robust; if it falls toward baseline levels, the shared BERT scorer inflated the result.","tokens_in":9662,"feed_emoji":"💬","tokens_out":3707,"duration_ms":36502,"temperature":0.7,"pith_summary":"The paper argues that generating a structured sentence—two clauses joined by a discourse marker—becomes easier if the model never mixes meaning and structure in one hidden vector. It defines a 'discourse transfer' task: given a head clause and a marker such as 'because' or 'but,' the model must produce a grammatical tail clause that preserves the stated relation. The proposed TransSent system encodes the head clause, translates its representation into a relation-specific space, and decodes the tail clause. Evaluations on three corpora and on free-text and dialogue tasks report higher relation accuracy and better human ratings than sequence-to-sequence, VAE, and planning baselines.","feed_headline":"Discourse markers become translation moves in sentence generation","feed_subtitle":"A new model splits meaning from structure, then translates a head clause into a tail clause that respects the marker.","key_machinery":"The load-bearing object is the relation translation network together with the two distance losses. In discourse space, head and tail clauses have vectors z_h and z_t; a learned matrix M_r projects them into a relation space, and the translated tail is W'[z_h M_r ⊕ z_r]. The L_dis loss minimizes the squared distance to the projected real tail, while L_ratio minimizes the ratio of that distance to the distance from the head, forcing the translation to point toward the tail rather than stay near the head. These losses, added to a reconstruction loss, form the training objective in Equation 10.","core_discovery":"The central claim is that a discourse marker between two clauses can be modeled as a translation operation on sentence embeddings, analogous to knowledge-graph translation models. TransSent's encoder-decoder reconstructs both head and tail clauses while a relation network projects clause representations into a relation-specific space and requires the translated tail representation to be close to the actual tail (L_dis) and farther from the head than from the tail (L_ratio). The training objective couples these losses with reconstruction. The paper claims that this explicit separation lets the model generate grammatical tails that preserve the intended relation, and that the same module transfers to free text and dialogue generation when paired with VAE or CVAE.","pith_inferences":["If the shared BERT scorer is biased toward TransSent's own outputs, the reported accuracy gains would shrink; an independent relation classifier trained without TransSent in the loop would settle this.","The translation analogy suggests that discourse markers could be composed recursively: chaining multiple markers would yield multi-clause sentences, which the paper itself lists as future work.","The ratio loss is a margin-like constraint that may generalize to other structured prediction tasks where a transformation should move a source representation toward a target rather than merely away from the source."],"forward_implications":["Structured sentences can be built compositionally: generate a head clause, pick a discourse marker, transfer to a tail clause, and concatenate the three parts.","The model can be combined with existing generation models such as VAE and CVAE, so discourse structure can be added to free-text and dialogue systems without retraining them from scratch.","Relation accuracy is notably lower on open-domain corpora than on a single-domain review corpus, suggesting that relation translation is harder when discourse content is diverse and pointing to larger or more targeted datasets as the next step.","The fine-tuned BERT used for DMP doubles as an automatic relation-accuracy scorer, so the same representation that guides generation also judges it."],"supporting_citations":[{"why":"Supplies the DMP task, the five discourse markers, the data-collection code, and the idea of extracting discourse pairs with dependency parsing.","marker":"Nie et al., 2017"},{"why":"Provides the BERT encoder that is fine-tuned on DMP and then used both as the sentence encoder and as the automatic relation-accuracy discriminator.","marker":"Devlin et al., 2018"},{"why":"TransR supplies the design of projecting entities and relations into separate spaces, which TransSent adapts to discourse and relation spaces.","marker":"Lin et al., 2015"},{"why":"TransE supplies the underlying translation hypothesis that the head representation plus a relation vector approximates the tail representation.","marker":"Bordes et al., 2013"},{"why":"CVAE is used as the head-response generator in the dialogue experiment, and its training data are used to fine-tune TransSent for dialogue.","marker":"Zhao et al., 2017"},{"why":"VAE is used as the head-discourse generator in the free text generation experiment.","marker":"Bowman et al., 2015"},{"why":"PHVM is the plan-based baseline adapted to discourse transfer by replacing attribute-value pairs with discourse markers.","marker":"Shao et al., 2019"}],"fun_headline_variants":["Treats discourse markers as translation between clauses","Translating head to tail via discourse markers","Split semantics and structure, translate via markers","TransSent: discourse markers as embedding translations","Generates structured sentences by translating clauses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The automatic relation-accuracy metric is computed by the same fine-tuned BERT that encodes the discourses, so if that BERT's relation space favors TransSent's own outputs, the reported accuracy gains could be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Treats discourse markers as translation between clauses","Translating head to tail via discourse markers","Split semantics and structure, translate via markers","TransSent: discourse markers as embedding translations","Generates structured sentences by translating clauses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1352,"prompt_tokens":861,"completion_tokens":491,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":426}},"tokens_in":477,"tokens_out":491,"duration_ms":5615,"temperature":1.0,"reasoning_tokens":426,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:50:34.370365+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train an independent discourse-marker classifier on the same DMP data but with a different architecture and no exposure to TransSent training, then run TransSent's generated tails through it. If its accuracy stays near the reported 50–80% on open-domain data, the metric is robust; if it falls toward baseline levels, the shared BERT scorer inflated the result.","supporting_citations":[{"cited_title":"Learning entity and relation embeddings for knowledge graph completion","cited_arxiv_id":null,"evidence_quote":"TransR supplies the design of projecting entities and relations into separate spaces, which TransSent adapts to discourse and relation spaces."},{"cited_title":"Translating embeddings for modeling multi-relational data","cited_arxiv_id":null,"evidence_quote":"TransE supplies the underlying translation hypothesis that the head representation plus a relation vector approximates the tail representation."}],"review_version":1}