Reading Between the Sentences: Discourse Relations as a Portable Signal for Implicit Toxicity

Much harmful text contains no harmful words. "They moved into our street last spring. The break-ins started in June." Each sentence is innocuous. The insult lives in the unstated link between them, the reader's inference that the second happened because of the first. Linguists call this an implicit discourse relation: a logical connection such as cause, contrast, or concession that the writer leaves out and the reader fills in. We propose to detect this kind of implicit toxicity by first recovering the missing connective, the "because" or "so" the author never wrote. We then score toxicity over the explicit readings the text makes available. The bet is that this link is more stable across genres and languages than the surface vocabulary current detectors learn. If so, making it explicit should shrink the performance drop detectors suffer when moved from, say, news comments to forum posts, or from English to a language they were never trained on.

The contribution has two parts. The first is a diagnostic evaluation: minimal pairs in which the two sentences stay fixed, only the implied relation changes, and toxicity flips. We build these across a grid of genres and languages, so we can measure whether a detector actually uses the relation. The second is a cheap intervention: a connective-prediction model whose distribution over possible connectives feeds a standard toxicity classifier. The diagnostic set is useful even if the intervention loses. We say below exactly which competitor would make the intervention unnecessary.

Core Claim

Toxicity that arises from an inferred relation between clauses should be scored by marginalizing over the readings a reader could plausibly construct. Write $a_1, a_2$ for two adjacent text segments and $c$ for a connective drawn from a fixed inventory (including "no relation"). Write $p_\theta(c \mid a_1, a_2)$ for a connective predictor's probability that a reader would insert $c$, and $q_\phi(\text{toxic} \mid a_1\,c\,a_2)$ for a toxicity classifier applied to the text with $c$ inserted. We score

$$ s(a_1,a_2) = \max_{c:\, p_\theta(c\mid a_1,a_2) > \tau} \; q_\phi(\text{toxic} \mid a_1\,c\,a_2), $$

and compare this against the expectation $\sum_c p_\theta(c \mid a_1,a_2)\, q_\phi(\text{toxic}\mid a_1\,c\,a_2)$. Here $\tau$ is a plausibility threshold. The max form encodes a specific claim: a dog whistle is harmful if any reasonably available reading is harmful, even a minority one.

This choice follows from DiscoGeM. That crowdsourced, genre-mixed corpus found that roughly 31% of implicit relations carry several valid readings at once. A single-label relation classifier would therefore throw away exactly the reading that deniable toxicity depends on. The portability claim is falsifiable. Detectors given relation-explicit input should lose less when moved across genres than detectors given raw text. A control that inserts random connectives should not produce the same benefit.

Why This Intersection

The four directions play different roles here. Implicit discourse relation recognition supplies the mechanism. Implicit toxicity is the application. Cross-genre and zero-shot cross-lingual transfer (training in one language and testing in another with no target-language labels) are the evaluation regimes.

Several strands of discourse work supply the method. Early work showed that meaning carried across sentence boundaries predicts the unstated relation. Marcu and colleagues did this with word pairs spanning the boundary, and work on the Penn Discourse Treebank and on automatic sense prediction established the task and its dependence on world knowledge. More recent systems predict the missing connective as an auxiliary target. These include auxiliary connective generation, connective prediction, and multi-token mask-filling. We reuse the mask-filling formulation directly because it outputs a distribution over surface connectives rather than a single sense label.

Shaik et al. train on explicit examples with the connective deleted. Ji et al. show that such artificially stripped text differs from naturally implicit text and treat the gap as domain adaptation. We take the Ji et al. result as a constraint. Our connective predictor must be validated on naturally implicit data, not only on stripped explicit data. All of this prior work stops at predicting the relation. None of it uses the recovered relation to judge what a text implies about a person or group, and none of it tests whether the relation transfers where lexical signals do not.

The cross-genre literature supplies the motivating evidence. Y'all should read this! argues that models exploit genre-specific shortcuts. Patterson and Kehler show that whether a writer drops a connective depends on register. News editing favors economy, and informal writing behaves differently. This bears directly on us, because it means the connective predictor itself may degrade across genres.

Faruqui and Padó report that an entailment model transfers better from English news to German news than from German news to German opinion journalism. Genre can matter more than language. That finding shapes our design: we evaluate on a genre-by-language grid rather than treating cross-lingual transfer alone. Liu et al. find that naively pooling reviews and blogs hurt sentiment classification. This is a negative result against the most obvious fix, and we test it rather than assume it.

On large language models, Lee et al. report that they favor topical similarity over logical structure. They notice that two sentences share a subject but miss whether one explains the other. If that holds, the problem matters for any approach that asks a language model to "explain the implication." It is also why an explicit relation step could help. Relation-based argument mining with LLMs and Stab and Gurevych show that support and attack relations without connectives are a recurring bottleneck downstream. Implicit toxicity is a close cousin: an unstated premise that supports a derogatory conclusion.

A caveat on the grounding. The research file available to us is truncated partway through the cross-genre section. It contains no indexed papers, benchmarks, or reported numbers for implicit toxicity detection or for zero-shot cross-lingual toxicity transfer. We therefore cannot confirm from the file that this exact combination is unexplored, and we cannot quote a baseline figure for it. The closest indexed work on transfer is zero-shot transfer for implicit discourse relation classification. It classifies relations from simple combinations of sentence embeddings (concatenation, difference, element-wise product) and transfers the relation classifier itself. It does not carry relations into a downstream judgment task.

Research Program

The first component is the connective predictor. We fine-tune a multilingual masked language model to fill a masked slot between $a_1$ and $a_2$ with connective tokens, following the multi-token mask-filling formulation. Training uses Penn Discourse Treebank implicit examples plus the genre-mixed DiscoGeM data. For the soft target on DiscoGeM, we use its crowd distributions over readings rather than the majority label.

For other languages, the model predicts connectives directly in the target language through its shared vocabulary. A small bilingual connective lexicon maps target-language connectives onto the shared sense inventory. This mapping is needed only for analysis, not for scoring. We measure the predictor's calibration per genre on DiscoGeM. Following Ji et al., we report separately on naturally implicit and on artificially stripped examples.

The second component is the toxicity scorer $q_\phi$. It is a multilingual encoder fine-tuned on source-genre, source-language implicit toxicity labels, with each training example expanded into its top-$k$ connective-inserted variants. The published implicit-toxicity corpora we would draw on are not indexed in our file. We will commit to specific ones when the protocol is registered, and we report no numbers for them here.

The third component is the diagnostic set, and we consider it the most durable contribution. Annotators write minimal pairs in which $a_1$ and $a_2$ are held fixed. Relation-cueing context outside the pair is then varied so that the preferred reading shifts, for example between a causal and a merely temporal reading. Toxicity changes with the reading. The pairs span at least three genres (news comments, forum threads, short social posts) and at least three languages. Native-speaker annotators write in-language pairs rather than translating them, since translation would import English genre conventions. Each item gets both a toxicity judgment and a DiscoGeM-style distribution over readings, so we can test whether toxicity tracks the minority reading.

Baselines

The naive reference point is a multilingual encoder fine-tuned on raw source-genre, source-language implicit toxicity data and applied unchanged to every other cell of the genre-by-language grid. It defines the transfer gap that everything else must close.

We found no reported result on Lacuna for any implicit toxicity detector under cross-genre or zero-shot cross-lingual transfer. Nor did we find one for connective-prediction methods on DiscoGeM or the Penn Discourse Treebank. The file names these methods but gives no figures. We will therefore reproduce all baselines ourselves rather than quote them.

The strongest competitor, in our judgment, is a large instruction-tuned language model. It is prompted to first write out, in free text, what the passage implies, and then to judge toxicity. The model and prompt are identical in every grid cell, and the model is given the same few source-language examples our system sees. This competitor performs the same "make the implicit explicit" move without committing to connectives. Lee et al.'s finding predicts it will be topic-biased, which makes it the right test of whether structure adds anything.

The second strong competitor is representation disentanglement in the style of Yuan et al.. Their method splits sentence representations into a context part and a genre-pattern part and discards the pattern part. We apply it to the toxicity encoder with genre as the nuisance variable, so it removes genre signal without adding any relation information.

We also run a pooling baseline that trains on all available genres in the source language. This directly tests Liu et al.'s warning against naive pooling.

Three ablations isolate the key ingredient. The first inserts connectives sampled uniformly at random, which controls for the effect of simply adding tokens and augmenting the data. The second inserts only the single argmax connective, which tests whether the multi-reading max matters. The third swaps the expectation for the max, which tests the deniability hypothesis.

All systems share the same training labels, encoder size (except the large-model competitor, which we report with its compute noted), and hyperparameter search budget. Tuning is done on source-cell development data only, never on target cells.

We count the proposal as beating a baseline when two conditions hold. First, the average relative drop from source cell to target cells, measured in macro-F1, must be smaller than the baseline's. Second, accuracy on the relation-flip minimal pairs must be higher. Both must hold with bootstrap confidence intervals that exclude zero. Winning on in-distribution accuracy alone does not count.

Why not just…?

Why not just prompt a large language model to explain the implication? This is the real threat, and we treat it as one. Lee et al. predict such models will latch onto the shared topic ("neighbors" and "crime") without committing to the causal link. Topic and harm are correlated in most data, so that bias could still yield good scores on ordinary test sets. The minimal pairs decide the question, because there the topic is constant and only the relation varies. If the prompted model matches our system on the minimal pairs across the grid, the intervention adds nothing. We would then report the diagnostic set and the genre-versus-language gap analysis as the contribution and drop the claim that connective prediction is needed.

Why not just pool training data from more genres? Liu et al. found that pooling mismatched genres hurt sentiment classification, but toxicity and modern encoders may behave differently. The pooling baseline answers this directly. The more informative test is held-out genres, since pooling can only cover genres one has labels for. The relation-based approach, if it works, should help on unseen genres too.

Why not just translate target-language text into English and run the English detector? Faruqui and Padó's result suggests the main obstacle is genre, not language, and translation does nothing about genre. Translation can also insert or drop connectives, silently changing the very relation we care about. We include translate-then-test as an extra baseline in the cross-lingual cells. If it closes the language gap but not the genre gap, that supports the grid framing. If it closes both, the cross-lingual half of our claim fails.

Why not just remove genre signal from the representation? Yuan et al.'s disentanglement discards nuisance style but supplies no relation reasoning. It is run as a baseline. Removing genre cues and adding relation structure are complementary, so we also report the two combined.

Proposed Experiments

The first experiment measures the grid. We train in one source cell (English news comments) and evaluate implicit-toxicity macro-F1 in every genre-by-language cell for all systems. The question is whether the genre gap exceeds the language gap for toxicity, as Faruqui and Padó found for entailment. Either answer is informative, and it decides where the method should focus.

The second experiment is the decisive check. We compare minimal-pair accuracy across the grid for the proposal, the prompted large language model, and the random-connective ablation. The mechanism is supported only if the proposal beats random insertion by a clear margin and does so in held-out genres, not just the source genre.

The third experiment tests the deniability hypothesis. On items whose toxic reading is a minority reading under the DiscoGeM-style annotation, the max score should beat both the expectation and the argmax, and the advantage should shrink on items where the toxic reading is the majority.

The fourth experiment is a dependency analysis. It correlates the connective predictor's per-genre calibration on DiscoGeM with the per-genre toxicity gain. If gains do not track predictor quality, the improvement is coming from somewhere other than the relation, and we will say so.