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.

Introduction

In natural language processing, understanding "who did what to whom" is the goal of Semantic Role Labeling (SRL). While many systems utilize PropBank for shallow, verb-specific roles, this work addresses the more ambitious FrameNet framework. FrameNet is built on the principle of frame semantics, which posits that word meanings are understood relative to conceptual scenes or "frames." For instance, words like buy and sell evoke a "Commerce" frame involving roles like Buyer and Seller.

Example of a frame-semantic parse for a complex sentence.

The complexity of FrameNet, which includes hundreds of frames and thousands of roles, presents a significant challenge, especially given the scarcity of annotated training data. Traditional local classifiers often fail to maintain global logical consistency, leading to errors such as overlapping roles or missing required participants.

Methodology: The SEMAFOR Parser

The proposed system, SEMAFOR, employs a three-stage pipeline to extract semantic structures:

  1. Target Identification: A rule-based component identifies words or phrases likely to evoke a frame.
  2. Frame Identification: This stage determines which specific frame a target evokes. To handle out-of-vocabulary words, the model uses a latent-variable log-linear approach. By representing words as "prototypes," the model can generalize from known words (e.g., waltz) to unknown ones (e.g., tango). Furthermore, graph-based semi-supervised learning is used to expand the lexicon by propagating frame information across word similarity graphs.
  3. Argument Identification: Once a frame is identified, the parser locates its arguments. Rather than predicting roles independently, SEMAFOR uses Dual Decomposition (specifically the AD3 algorithm) to treat this as a constrained optimization problem. This ensures the output respects linguistic constraints such as role uniqueness and mutual exclusivity.
Comparison of syntactic dependency structure (a) and frame-semantic structure (b).

Experimental Results

The system was evaluated on the SemEval 2007 (Task 19) benchmark and the larger FrameNet 1.5 dataset. Key findings include:

  • State-of-the-Art Accuracy: SEMAFOR outperformed previous systems on the SemEval 2007 benchmark, achieving a gain of approximately 3-4 F1 points in full-text parsing.
  • Computational Efficiency: The AD3 dual decomposition algorithm proved to be nine times faster than the industry-standard CPLEX solver while maintaining comparable accuracy.
  • Generalization: The semi-supervised approach improved frame identification for unknown words by nearly 20% over a purely supervised baseline.

Discussion and Limitations

While SEMAFOR represents a significant advancement, several limitations remain. The parser is highly dependent on the accuracy of the external syntactic dependency parser; errors in syntax often propagate to the semantic level. Additionally, the current model is limited to intra-sentential arguments and struggles with discontinuous multi-word expressions. Future work may address cross-sentence role filling to better capture discourse-level semantics.

SEMAFOR has been released as open-source software to facilitate further research in downstream applications such as question answering and information extraction.

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