Relational triple extraction pulls facts of the form (subject, relation, object) out of sentences, for example (Beijing, capital_of, China). One strong family of models writes every pair of tokens into a square table and marks each triple as a rectangle whose rows cover the subject and whose columns cover the object. These models already search in two directions: they decode from subject to object and from object to subject, so that a subject that is hard to spot can be recovered through its object. Even so, a large share of their remaining mistakes are not missed entities. They are swapped roles. On WebNLG, 45.1% of the errors of a strong span-linking model had the right entities and the right relation but put head and tail the wrong way round (Relational Triple Extraction: One Step is Enough). A second large share of errors are boundary errors on multi-token entities.
We propose to treat the two search directions as two readings of the same table: reading cell $(i,j)$ as "$i$ is subject, $j$ is object," and cell $(j,i)$ as the reverse. The table is processed by a convolution stack with several dilation rates. A dilated convolution is a filter whose taps are spaced apart, so it sees a wider neighborhood without adding weights. The output is then split exactly into a part that is symmetric under transposing the table and a part that flips sign. The symmetric part decides whether two spans are related at all. The sign-flipping part decides which span is the head. Swapping roles is therefore no longer a separate mistake the model can make independently of detection: by construction, evidence for "$i$ heads $j$" is evidence against "$j$ heads $i$." The dilation rates are set separately for rows and columns, so that the filter centered on one corner of a triple's rectangle can reach the opposite corner for subjects and objects of different lengths. That reach is what boundary precision on long entities requires.
Core Claim
Dual search in current models is procedural. BiRTE runs two extraction streams over a shared encoder and filters the merged candidates with a biaffine classifier. OD-RTE walks the token-pair grid from two opposite diagonal corners. SBN and B-MRC pair aspect-to-opinion and opinion-to-aspect decoders or questions. In all of these, the two directions are reconciled after the fact by confidence rules. Nothing prevents both directions from firing confidently on the same pair.
Our claim is that making the two readings algebraic, as the sum and difference of one table and its transpose, removes a large fraction of role inversions without hurting detection. We further claim that anisotropic multi-scale dilation on that table improves exact-match boundaries on long entities beyond what a parameter-matched stack of plain $3\times 3$ convolutions achieves. These are two separable claims, and the experiments are designed so that either can fail on its own.
The closest prior systems each have one ingredient and lack the other. OD-RTE has vertex tagging and bidirectional decoding on the grid but no spatial convolution. BDTF runs ResNet-style $3\times 3$ convolutions over an aspect–opinion table, and loses about 8.7% when they are removed. That result motivates spatial modeling of the table, but BDTF's decoding is a one-direction detect-then-classify pipeline with no dilation. Neither system ties the $(i,j)$ and $(j,i)$ cells together.
Research Program
Let $h_1,\dots,h_L$ be encoder states for a sentence of $L$ tokens. We form a cell table $H_{ij} = \phi([h_i; h_j; h_i \odot h_j])$, which is deliberately not symmetric in $i,j$. We pass it through a convolution stack $G$ and write $Z = G(H)$. We then define
$$S = \tfrac{1}{2}(Z + Z^\top), \qquad A = \tfrac{1}{2}(Z - Z^\top),$$
where the transpose swaps the two token indices of every cell, so that $S_{ij} = S_{ji}$ and $A_{ij} = -A_{ji}$. For relation $r$, the score that span-cell $i$ is head and $j$ is tail is
$$\ell_r(i \to j) = u_r^\top S_{ij} + v_r^\top A_{ij},$$
with learned vectors $u_r, v_r$. It follows automatically that $\ell_r(j \to i) = u_r^\top S_{ij} - v_r^\top A_{ij}$. The detection evidence is shared, and the orientation evidence enters with opposite signs. For symmetric relations such as spouse, $v_r$ can shrink toward zero, and both directions are emitted.
The split also fits OD-RTE's four-corner vertex labels cleanly. Under transposition, a box's upper-left and lower-right corners map to the same labels of the reversed box, while upper-right and lower-left swap. The upper-left and lower-right heads therefore read $S$ and $A$ directly, and the upper-right and lower-left heads are tied as a swapped pair. Decoding keeps OD-RTE's bidirectional diagonal walk and relation negative sampling, so any gain is attributable to the representation, not to a new decoder.
One case needs care. On the diagonal, $A_{ii}=0$, so a cell where subject and object start at the same token carries no orientation signal. Subject–object overlap cases whose other corners lie off the diagonal still get a signal. We will report the subject–object overlap subset separately so that any weakness here is visible.
The convolution stack $G$ reuses a known method, factorized asymmetric kernels from CFPNet. A $3\times1$ filter along rows and a $1\times3$ filter along columns are each given their own dilation rate. Channel groups get different (row, column) dilation pairs drawn from $\{1,2,4\}^2$, and CFPNet's hierarchical feature fusion sums the groups step by step. This summation is the device CFPNet uses to suppress gridding, the checkerboard gaps left when a sparse filter skips neighboring cells. We use it because a single high dilation on a text table would skip exactly the adjacent cells that make up a contiguous entity such as "New York City."
The motivation for separate row and column rates is geometric. A triple's box has height equal to the subject length and width equal to the object length. A cell at one corner can see the opposite corner only if the filter reaches that offset along each axis, and subjects and objects are often of very different length. PSConv's cyclic channel lattice and SCSC's content-gated fusion of kernel sizes are alternative multi-scale operators. We will test them as variants of $G$, but our default is the factorized form because it keeps the row and column axes, and therefore the subject and object roles, explicitly separate.
On the vision side, the evidence that parallel dilations with dense fill-in beat uniform kernels comes from segmentation, detection and dehazing (CFPNet, PSConv, HRGAN). That evidence motivates the design but does not transfer by default: text tables are small, and most entities are short. This is why the plain-convolution control below is central rather than a formality.
Baselines
The naive reference point is OD-RTE as published, retrained by us on the same encoder, splits and seeds. The file gives no F1 for it: no reported result was found on Lacuna for OD-RTE's F1 on NYT or WebNLG. The file does give its efficiency on NYT, 8.3 GB of GPU memory and 798 seconds per epoch. Our model must stay within a small factor of those figures to count as a practical improvement.
The strongest competitor with a reported number is BiRTE (A Simple but Effective Bidirectional Framework). It reports 92.8 exact-match F1 on NYT and 89.3 exact-match F1 on WebNLG, and it is the clearest embodiment of procedural dual search. DirectRel reports 90.0 F1 on WebNLG, and its error analysis supplies the role-error target we aim to reduce. GRTE represents the attention-based alternative for mixing table cells. The file gives only relative gains for it, +1.9% on NYT29 and +3.3% on WebNLG, so we will rerun it.
All systems will use the same BERT-base encoder, the same data splits and a matched hyperparameter search budget (the same number of trials per system). Published numbers are quoted alongside our reruns rather than substituted for them.
Beating BiRTE means exceeding 89.3 exact-match F1 on WebNLG and 92.8 on NYT, averaged over at least three seeds. On WebNLG we additionally require a clear drop in the share of errors that are role inversions, relative to both our OD-RTE rerun and the 45.1% DirectRel figure.
Four ablations isolate the ingredients. First, the transpose split applied directly to $H$ with no convolution tests whether the algebra alone fixes roles. Second, the full convolution stack without the split, scoring cells directly from $Z$, tests whether convolution alone fixes roles. Third, a stack of plain $3\times 3$ convolutions in the BDTF style, matched to our model in both parameter count and receptive field, isolates what multi-scale dilation contributes. Fourth, a single uniform dilation rate checks the gridding concern.
Why not just…?
Why not just add BDTF-style plain convolutions to OD-RTE? This is the real threat to the multi-scale half of the proposal. Most entities span only a few tokens, so a few $3\times 3$ layers may already cover the relevant corner offsets. The matched plain-convolution ablation settles the question. If it equals the anisotropic dilated stack on long-entity exact match, we drop the multi-scale claim and report the transpose split as the contribution, since the split does not depend on which convolution is used.
Why not just give each cell a relative-position or "which token comes first" feature? Word order is not role: in "China's capital, Beijing," the object precedes the subject. A position feature can therefore learn only a dataset-level prior about order. We include this as a baseline and expect it to reduce role errors only where order and role coincide. The comparison stratified by head-first versus tail-first triples will show whether that happens.
Why not just add a consistency loss between the $s\to o$ and $o\to s$ streams, as in BiRTE plus a penalty? A soft penalty encourages what the split enforces exactly, at the cost of an extra weight to tune and two separate scoring paths. This is a fair alternative, and we will implement it on our backbone with the same budget. If the soft version matches the hard split on role errors, the contribution reduces to the convolutional encoder and we will say so.
Why not just rely on the encoder's self-attention, which already sees all tokens? BDTF's 8.7% drop without table convolutions, and GRTE's need for extra rounds of table-level attention, both indicate that token-level encodings do not supply cell-to-cell structure on their own. The GRTE rerun also tests attention over cells against convolution at matched compute.
Proposed Experiments
The main benchmarks are NYT and WebNLG under exact match, together with their single-entity-overlap, entity-pair-overlap and subject–object-overlap subsets. As a transfer check of the convolutional component alone, we will also use the ASTE sets 14LAP, 14RES, 15RES and 16RES, where the anisotropic dilated stack replaces BDTF's CNN inside BDTF itself. Aspects and opinions are different types, so the transpose split is not applied there.
Beyond F1, we report three diagnostics. The first is the role-inversion rate: predictions $(e_2,r,e_1)$ where only $(e_1,r,e_2)$ is gold, following DirectRel's error categories. The second is exact-match F1 bucketed by subject length and object length. The third is the gap between exact-match and partial-match F1, which isolates boundary errors.
Three results are decisive. The transpose split should cut the WebNLG role-inversion rate substantially relative to the no-split ablation at unchanged recall of unordered pairs. If recall drops, the shared detection term is being hurt by the sign constraint. The anisotropic dilated stack should beat the matched plain stack specifically in the buckets where the subject and object lengths differ, since that is where separate row and column reach should matter. If the gain is uniform across buckets or absent, the geometric motivation is wrong. Finally, the full model should exceed BiRTE's reported 89.3 WebNLG and 92.8 NYT exact-match F1 while staying near OD-RTE's 8.3 GB and 798-second-per-epoch budget on NYT.
References
- Relational Triple Extraction: One Step is Enough
- A Simple but Effective Bidirectional Framework for Relational Triple Extraction
- OD-RTE: A One-Stage Object Detection Framework for Relational Triple Extraction
- A Span-level Bidirectional Network for Aspect Sentiment Triplet Extraction
- Bidirectional Machine Reading Comprehension for Aspect Sentiment Triplet Extraction
- Boundary-Driven Table-Filling for Aspect Sentiment Triplet Extraction
- CFPNET: Channel-Wise Feature Pyramid For Real-Time Semantic Segmentation
- PSConv: Squeezing Feature Pyramid into One Compact Poly-Scale Convolutional Layer
- SCSC: Spatial Cross-scale Convolution Module to Strengthen both CNNs and Transformers
- Visual Haze Removal by a Unified Generative Adversarial Network
- A Novel Global Feature-Oriented Relational Triple Extraction Model based on Table Filling