Frame-Semantic Parsing

Dipanjan Das, Desai Chen, André F. T. Martins, Nathan Schneider, Noah A. Smith

Abstract

Frame semantics is a linguistic theory that has been instantiated for English in the FrameNet lexicon. We solve the problem of frame-semantic parsing using a two-stage statistical model that takes lexical targets (i.e., content words and phrases) in their sentential contexts and predicts frame-semantic structures. Given a target in context, the first stage disambiguates it to a semantic frame. This model uses latent variables and semi-supervised learning to improve frame disambiguation for targets unseen at training time. The second stage finds the target's locally expressed semantic arguments. At inference time, a fast exact dual decomposition algorithm collectively predicts all the arguments of a frame at once in order to respect declaratively stated linguistic constraints, resulting in qualitatively better structures than naïve local predictors. Both components are feature-based and discriminatively trained on a small set of annotated frame-semantic parses. On the SemEval 2007 benchmark data set, the approach, along with a heuristic identifier of frame-evoking targets, outperforms the prior state of the art by significant margins. Additionally, we present experiments on the much larger FrameNet 1.5 data set. We have released our frame-semantic parser as open-source software.

Introduction: Beyond Verb-Centric Semantics

Traditional Semantic Role Labeling (SRL) often relies on the PropBank style, which is closely tied to verbs and syntactic structures. In contrast, frame semantics, as instantiated in the FrameNet lexicon, offers a more comprehensive view by defining "frames"—conceptual structures like Self_motion or Collaboration—that can be evoked by a wide variety of lexical units, including nouns, adjectives, and prepositions.

Comparison between PropBank-style SRL (a) and FrameNet-style parsing (b).

The complexity of FrameNet parsing arises from two main areas: lexical sparsity and structural consistency. FrameNet is vast, but training data is limited, making it difficult to identify frames for unseen words. Furthermore, frames often have multiple roles (e.g., Agent, Goal) that must be assigned to word spans without violating linguistic constraints, such as role uniqueness or mutual exclusion.

The SEMAFOR Parser

The authors introduce SEMAFOR, a high-performance parser that addresses these challenges through a two-stage statistical model.

Frame Identification with Latent Variables

To handle words not seen during training, SEMAFOR uses a latent-variable model. Instead of a direct mapping, the model relates target words to "prototype" lexical units known to evoke specific frames. This is further bolstered by graph-based semi-supervised learning (SSL), which propagates frame labels across a semantic similarity graph of millions of words.

Joint Argument Identification via Dual Decomposition

The second stage identifies semantic arguments for the chosen frame. Rather than treating each role independently—which often leads to overlapping or contradictory assignments—the authors formulate this as a constrained optimization problem.

An example sentence annotated with FrameNet frames and roles.

Using the Alternating Directions Dual Decomposition (AD3) algorithm, the parser finds a globally optimal assignment of roles to spans. This ensures that the resulting structure respects declaratively stated linguistic constraints, such as ensuring a single span does not fill two incompatible roles.

Results and Performance

SEMAFOR was evaluated on the SemEval 2007 benchmark and the larger FrameNet 1.5 dataset. Key findings include:

  • State-of-the-Art Accuracy: SEMAFOR outperformed previous systems by over 4.4 F1 points on the full parsing task.
  • Generalization: The SSL-based lexicon expansion improved frame identification for unknown words by approximately 20%.
  • Computational Efficiency: The AD3 decoder proved to be 9 times faster than standard CPLEX ILP solvers while achieving identical accuracy.

Conclusion

By combining latent-variable modeling for lexical coverage with dual decomposition for structural consistency, SEMAFOR provides a robust and efficient solution for frame-semantic parsing. The system's ability to respect global constraints while remaining computationally viable marks a significant advancement in deep semantic analysis.

Summary updated

Abstract

Frame semantics is a linguistic theory that has been instantiated for English in the FrameNet lexicon. We solve the problem of frame-semantic parsing using a two-stage statistical model that takes lexical targets (i.e., content words and phrases) in their sentential contexts and predicts frame-semantic structures. Given a target in context, the first stage disambiguates it to a semantic frame. This model uses latent variables and semi-supervised learning to improve frame disambiguation for targets unseen at training time. The second stage finds the target's locally expressed semantic arguments. At inference time, a fast exact dual decomposition algorithm collectively predicts all the arguments of a frame at once in order to respect declaratively stated linguistic constraints, resulting in qualitatively better structures than naïve local predictors. Both components are feature-based and discriminatively trained on a small set of annotated frame-semantic parses. On the SemEval 2007 benchmark data set, the approach, along with a heuristic identifier of frame-evoking targets, outperforms the prior state of the art by significant margins. Additionally, we present experiments on the much larger FrameNet 1.5 data set. We have released our frame-semantic parser as open-source software.
BibTeX
@inproceedings{das2014frame,
  author = {Dipanjan Das and Desai Chen and André F. T. Martins and Nathan Schneider and Noah A. Smith},
  title = {{Frame-Semantic Parsing}},
  booktitle = {Comput. Linguistics},
  year = {2014},
  url = {https://openreview.net/forum?id=wHhU4NJrqv5},
}
Figure