AMR-EnsembleNet: Consensus Graph Parsing

• AMR-EnsembleNet is a graph ensembling framework that aggregates heterogeneous AMR parser outputs using a pivot-and-vote strategy to form a consensus graph.
• It formalizes ensemble graph prediction as an NP-hard optimization problem, employing heuristic approximations and best-match mappings between graph components.
• Experiments demonstrate state-of-the-art gains with ensemble Smatch scores improving from 84.22 to 86.26 on AMR benchmarks and enhanced robustness on out-of-distribution data.

Searching arXiv for the specified paper and closely related AMR parsing context. AMR-EnsembleNet may be understood as the AMR-specific graph ensembling framework introduced under the heuristic name “Graphene” in “Ensembling Graph Predictions for AMR Parsing.” It addresses the problem of combining multiple AMR parser outputs into a single consensus graph by formalizing ensemble graph prediction as mining the largest graph that is most supported by a collection of graph predictions. The work treats the underlying optimization as NP-hard, proposes an efficient heuristic approximation, and evaluates it on five standard AMR benchmarks, where the ensemble predictions are reported to be more accurate than any individual models (Lam et al., 2021).

1. Problem setting and conceptual scope

The framework is situated in the broader class of machine-learning tasks in which models predict structured objects rather than scalar labels. The motivating examples given are dependency trees and Abstract Meaning Representation (AMR) graphs. Whereas ensemble methods are well established for classification and regression, the paper identifies ensemble graph prediction as comparatively underexplored and formulates it as a distinct combinatorial problem (Lam et al., 2021).

A possible misconception is that graph ensembling is a straightforward majority vote over local decisions. The formulation here is more restrictive: support must be defined relative to graph-to-graph alignments, because vertices and edges in different predicted graphs are not natively co-indexed. This makes the ensemble problem dependent on best-match mappings between graphs and leads directly to a nontrivial optimization problem over directed, labeled graph structure.

Within AMR parsing, the framework is designed to aggregate the complementary error profiles of heterogeneous parsers. The reported system combines outputs from sequence-to-sequence, transition-based, and graph-based incremental parsers, along with multiple seeds of a single architecture, and then selects a consensus graph by support maximization rather than by choosing a single parser output a priori (Lam et al., 2021).

2. Formalization of ensemble graph prediction

Let $G=\{g_1,\dots,g_m\}$ be $m$ directed, labeled graphs, each $g_i=(V_i,E_i)$ with vertex-label function $\ell(v)$ and edge-label function $\ell(e)$. The objective is to construct a new graph $g=(V,E)$ whose vertices and edges are well supported by the collection $G$ (Lam et al., 2021).

Support is defined through a best-matching $\phi_i=\phi(g,g_i)$, which is a bijection between a subset of $V$ and a subset of $V_i$. Under $m$0, each vertex contributes

$m$1

and each edge $m$2 contributes

$m$3

The total support accumulated over the ensemble is then

$m$4

Given a threshold $m$5, typically $m$6, a graph $m$7 is $m$8-supported if every vertex and every edge in $m$9 satisfies the threshold condition. The target graph is defined as

$g_i=(V_i,E_i)$0

This objective makes explicit that the ensemble is not merely selecting an existing prediction; it is searching for a graph whose local components are globally well supported under induced graph alignments. In AMR terms, this amounts to consensus over concepts and relations after approximate structural matching, rather than over linearized decoding tokens or parser confidences.

3. Computational hardness and the Graphene heuristic

The optimization is NP-hard. The hardness sketch proceeds by considering the special case $g_i=(V_i,E_i)$1, with all labels identical and $g_i=(V_i,E_i)$2. Under these constraints, the ensemble graph must be a common subgraph of $g_i=(V_i,E_i)$3 and $g_i=(V_i,E_i)$4, and maximizing total support becomes the maximum-edge common subgraph (MCES) problem, which is stated to be NP-complete. Consequently, even deciding an optimal ensemble graph is NP-hard (Lam et al., 2021).

The proposed heuristic, “Graphene,” avoids exhaustive search over all $g_i=(V_i,E_i)$5-supported graphs and instead uses a pivot-and-correct strategy. For each model index $g_i=(V_i,E_i)$6, the graph $g_i=(V_i,E_i)$7 is treated as a pivot graph. A vote table $g_i=(V_i,E_i)$8 is initialized over existing pivot vertices and edges, with each pivot node and edge receiving one vote from itself. For every other graph $g_i=(V_i,E_i)$9, the algorithm computes a best match $\ell(v)$0, for example using Smatch’s hill-climbing, and updates votes for node labels and edge labels according to the aligned structures. After aggregating votes from the remaining $\ell(v)$1 graphs, a filtering step retains only labels and relations whose counts meet the threshold $\ell(v)$2.

The procedure can be summarized as follows: each pivot produces a candidate corrected graph $\ell(v)$3; the final ensemble output is the candidate with the largest total support. In AMR parsing, the method enforces that each node and edge has exactly one label, so ties are broken by maximum vote, or arbitrarily if still tied. The description also notes that disconnected edges below $\ell(v)$4 can be optionally retained if one wishes to maintain connectivity (Lam et al., 2021).

The key data structures are a vote table of size $\ell(v)$5 for node-label candidates and $\ell(v)$6 for edge-label candidates, where $\ell(v)$7 and $\ell(v)$8 are small sets of labels observed across the ensemble. BestMatch is reused from the Smatch implementation and is reported to run in practice in $\ell(v)$9 with $\ell(e)$0 hill-climb restarts. If $\ell(e)$1 denotes average graph size and $\ell(e)$2 the number of models, each pivot requires $\ell(e)$3 matchings at cost $\ell(e)$4 plus $\ell(e)$5 per-vote updates, yielding total complexity $\ell(e)$6. The practical regime reported is $\ell(e)$7, $\ell(e)$8, with each ensemble run taking less than $\ell(e)$9 seconds on $g=(V,E)$0 CPU cores (Lam et al., 2021).

Two approximations are identified explicitly. The first is the use of greedy hill-climbing to approximate BestMatch. The second is the restriction of candidate graphs to corrections of pivots rather than arbitrary node and edge subsets. This suggests that the method trades exact search optimality for a constrained but computationally manageable consensus space.

4. AMR-specific instantiation and experimental design

The AMR instantiation ensembles four state-of-the-art parsers plus three additional checkpoints of SPRING, for a total of $g=(V,E)$1 models: SPRING (seq2seq BART-based) with four seeds, T5 (seq2seq on linearized AMR), APT (transition-based Transformer), and Cai & Lam (graph-based incremental) (Lam et al., 2021).

The voting scheme is parser-output centric. Each pivot node or edge receives one vote for itself and one vote per matched label or edge-label from each other parser under the best-match mapping $g=(V,E)$2. The threshold is set to $g=(V,E)$3 to mimic majority vote. After voting, only labels and edges with at least $g=(V,E)$4 votes are kept, with multi-label ties resolved by highest count (Lam et al., 2021).

The reported datasets are:

• AMR 2.0: $g=(V,E)$5 train, $g=(V,E)$6 dev, $g=(V,E)$7 test.
• AMR 3.0: $g=(V,E)$8 train, $g=(V,E)$9 dev, $G$0 test.
• BIO: $G$1 test, out-of-distribution, trained on AMR 2.0.
• Little Prince (LP): $G$2 test, out-of-distribution.
• New3: $G$3 test, out-of-distribution.

The baseline parsers are SPRING, T5, APT, and Cai & Lam. For SPRING, the official checkpoint plus three re-trained seeds are used. T5 is trained for $G$4 epochs with Adam, learning rate $G$5, batch size $G$6, and a small auxiliary word-sense task. APT and Cai & Lam use the authors’ best settings, with APT augmented by $G$7 silver data (Lam et al., 2021).

Evaluation uses Smatch overall and the following subscores: Unlabeled, No-WSD, Concepts, Named Entities, Wikification, Negations, Reentrancies, and SRL ($G$8 roles). The implementation is in Python, reusing Smatch code for BestMatch. Ensemble execution is on an $G$9-core CPU, and parser decoding is on a V100 GPU (Lam et al., 2021).

5. Reported performance across benchmarks

On in-distribution benchmarks, the reported Smatch scores are as follows. For AMR 2.0, SPRING (best) attains $\phi_i=\phi(g,g_i)$0, APT $\phi_i=\phi(g,g_i)$1, T5 $\phi_i=\phi(g,g_i)$2, and Cai & Lam $\phi_i=\phi(g,g_i)$3. Graphene with four SPRING models reaches $\phi_i=\phi(g,g_i)$4, and Graphene with all seven models reaches $\phi_i=\phi(g,g_i)$5. The previous S-OTA listed in the exposition is $\phi_i=\phi(g,g_i)$6. For AMR 3.0, SPRING (best) attains $\phi_i=\phi(g,g_i)$7, APT $\phi_i=\phi(g,g_i)$8, T5 $\phi_i=\phi(g,g_i)$9, Graphene (4 SPRING) $V$0, Graphene (all 7) $V$1, and previous S-OTA $V$2 (Lam et al., 2021).

The paper further reports that all submetrics rise under the full ensemble. Two examples given explicitly are Named-Entity $V$3, which increases from $V$4 to $V$5, and Reentrancy, which increases from $V$6 to $V$7. These results indicate that the gains are not restricted to the aggregate Smatch score but also extend to structurally and semantically salient AMR subproblems (Lam et al., 2021).

On out-of-distribution data, the reported scores are likewise improved by ensembling. For BIO, the baseline parser scores are SPRING $V$8, APT $V$9, T5 $V_i$0, and Cai & Lam $V_i$1; Graphene (4S) reaches $V_i$2, Graphene (all) reaches $V_i$3, and previous S-OTA is $V_i$4. For New3, the corresponding figures are $V_i$5, $V_i$6, $V_i$7, $V_i$8, $V_i$9, $m$00, and $m$01. For LP, they are $m$02, $m$03, $m$04, $m$05, $m$06, $m$07, and $m$08 (Lam et al., 2021).

The exposition characterizes these out-of-distribution outcomes as new state of the art: BIO $m$09, New3 $m$10, and LP $m$11. A plausible implication is that graph-level consensus over heterogeneous parser outputs remains beneficial even under substantial domain shift, rather than merely regularizing in-domain variance.

6. Ablations, selection criteria, and interpretive implications

The ablation analysis addresses which components of the framework drive the observed gains. With respect to pivot source, the appendix is said to show that all seven pivots contribute and that no single parser dominates. In BIO, described as the largest-shift setting, votes from T5 and APT are reported to be particularly important (Lam et al., 2021).

The threshold parameter is tuned on the AMR 2.0 development set, and $m$12 in majority-vote terms is reported to balance precision and recall best. This is consistent with the implementation choice $m$13 for the seven-model ensemble and situates the method between permissive union-like aggregation and high-precision intersection-like aggregation (Lam et al., 2021).

Candidate selection is examined through two alternatives: Graphene_support, which selects the candidate with largest total support, and Graphene_smatch, which selects by highest average Smatch among candidates, following Barzdins & Gosko. The support-based criterion is described as model-agnostic, while the Smatch-based criterion gives a small extra $m$14–$m$15 in-distribution but fails on one out-of-distribution set. The normalized total-support score is reported to correlate with Smatch at Pearson $m$16 with $m$17, which is presented as justification for optimizing support as a proxy (Lam et al., 2021).

The robustness analysis under reduced training data reports that when training on $m$18 or $m$19 of AMR 2.0, Graphene’s gains over the best single model increase, with one explicit example being an increase from $m$20 to $m$21 on AMR 2.0. This suggests that ensemble graph prediction is particularly useful in lower-resource conditions, where model-specific error modes are less fully corrected by training data alone (Lam et al., 2021).

Taken together, the ablations identify three recurring factors: majority-vote thresholding, support maximization, and pivot diversification. In the reported interpretation, these components enable the pivot-and-vote heuristic to combine complementary errors of modern AMR parsers into a consensus graph that improves over every individual system both in-domain and under domain shift.