Research map / Probabilistic and scientific modeling / Molecules and proteins
Frame Semantic Role Labeling Using Arbitrary-Order Conditional Random Fields
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes arbitrary-order CRFs with CPD-decomposed full factors and mean-field variational inference unrolled as an RNN, on top of a BERT cross-encoder, for end-to-end frame semantic role labeling on FrameNet 1.5/1.7.
Problem
Frame semantic role labeling (FSRL) identifies and labels arguments of frame-evoking targets with frame elements, but prior methods do not adequately model high-order interactions among arguments.
Approach
Learn a per-frame CRF with one variable per frame element and a single full factor over all elements; make it tractable by CPD (Kruskal form) of the factor tensor and approximate it with mean-field variational inference unrolled into a recurrent network, combined with span-based scoring and decoding.
Key results
- On FN1.5 without exemplars, arbitrary-order model reaches 74.67 F1 vs. AGED reproduction 74.15 and first-order 74.42.
- On FN1.7 with exemplars, arbitrary-order model reaches 77.06 F1 vs. AGED reproduction 76.62 and reported 76.84.
- Ablation on FN1.7 shows span-based scoring gives the largest gain among first-order changes (74.84 to 75.28 combined with other changes).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use arbitrary-order inference instead of first-order inference | FSRL model · FN1.5 test, w/o exemplar | F1-score: 74.42 -> 74.67 | Helped |
| Use arbitrary-order inference instead of first-order inference | FSRL model · FN1.7 test, w/o exemplar | F1-score: 75.31 -> 75.46 | Helped |
| Use arbitrary-order inference instead of first-order inference | FSRL model · FN1.5 test, w/ exemplar | F1-score: 76.68 -> 76.83 | Helped |
| Use arbitrary-order inference instead of first-order inference | FSRL model · FN1.7 test, w/ exemplar | F1-score: 76.90 -> 77.06 | Helped |
| Add MLPs (ablation) | AGED · FN1.7 without exemplar | F1-score: 74.76 -> 74.84 | Helped |
| Add gradient clipping (ablation) | AGED w/ MLPs · FN1.7 without exemplar | F1-score: 74.84 -> 74.94 | Helped |
| Set unary scores of null spans to 0 (ablation) | AGED w/ MLPs & GC · FN1.7 without exemplar | F1-score: 75.09 -> 75.17 | Helped |
| Use span-based method instead of pointer network (ablation) | AGED w/ MLPs & GC · FN1.7 without exemplar | F1-score: 75.09 -> 75.28 | Helped |
| Combine MLPs, gradient clipping, span-based method, and Zero setting (ablation) | AGED · FN1.7 without exemplar | F1-score: 75.28 -> 75.31 | Helped |
Surprising findings
- The authors could not reproduce the AGED official code results and had to add MLPs, gradient clipping and tuning to match or exceed them.
Limitations
- Gains over the first-order model are modest (about 0.2-0.4 F1 on the test sets)
- Assumes a single argument per frame element (multiple arguments sharing a role are excluded)
- Requires learning a separate CRF per frame, sharing only the encoder and scorer
Experimental setup
Datasets: FrameNet 1.5 FrameNet 1.7 FrameNet exemplar sentences
Benchmarks: FrameNet 1.5 FrameNet 1.7
Metrics: micro-F1 (exact span match) precision recall
Baselines: semi-CRF Lin, Sun, and Zhang (2021) Kalyanpur et al. (2020) AGED Chen, Zheng, and Chang (2021) Bastianelli, Vanzo, and Lemon (2020) KID
Builds on: AGED CRF as recurrent neural networks pointer networks regularized Frank-Wolfe
Closest papers by meaning
- Span-Based Semantic Role Labeling with Argument Pruning and Second-Order Inference (AAAI 2022 · similarity 0.90)
- Query Your Model with Definitions in FrameNet: An Effective Method for Frame Semantic Role Labeling (AAAI 2023 · similarity 0.88)
- Syntax-Aware Neural Semantic Role Labeling (AAAI 2019 · similarity 0.88)
- Short-Context Dominance: How Much Local Context Natural Language Actually Needs? (NeurIPS 2026 · similarity 0.85)
- Towards Auto-Regressive Next-Token Prediction: In-context Learning Emerges from Generalization (ICLR 2025 · similarity 0.83)
- Causal Graphical Models for Vision-Language Compositional Understanding (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.