Identifying Physical Object Use in Sentences

Tianyu Jiang, Ellen RiloffEMNLP 2022 · DOI

Abstract

Determining how physical objects are utilized within a narrative is a fundamental component of human commonsense reasoning. We introduce Object Use Classification (OUC), a three-way classification task to identify if an object mentioned in text is currently being used, is intended for future use, or is not being used. We present the TOUCAN (Textual Object Use ClAssificatioN) model, which integrates RoBERTa with external knowledge from FrameNet and data augmentation techniques. Our model achieves 72.4% accuracy on a newly curated dataset of 2,123 sentences, significantly outperforming large language model baselines.

Introduction

Humans naturally infer implicit actions from text; for instance, reading "Sam enjoyed the book" immediately implies the action of reading. While previous NLP research has mapped objects to their prototypical functions (e.g., knives for cutting), a significant challenge remains: objects are often mentioned in contexts where they are passive or unused. An AI that assumes every mentioned object is being used for its primary function will frequently misinterpret narratives. To address this, we propose that recognizing if an object is being used is a necessary precursor to determining how it is used.

The Object Use Classification Task

We define OUC as a three-way classification problem:

  1. Used: The object is actively involved in an action or a result.
  2. Anticipated Use: The text implies an intent to use the object in the near future.
  3. No Use: The object is mentioned but remains passive or is being stored.
Examples of sentences categorized by object use and their corresponding prototypical functions.

Our analysis of the TOUCAN dataset reveals that when an artifact is used, it follows its prototypical function 96% of the time. However, nearly 30% of use cases involve implicit actions not explicitly stated by a verb, highlighting the necessity of the OUC task for deep narrative understanding.

Methodology: The TOUCAN Model

The TOUCAN model is built upon a RoBERTa-base architecture and enhanced with three primary strategies to improve generalization and context-awareness:

  • Data Augmentation: We utilize WordNet for synonym and hyponym replacement and employ back-translation (English to Chinese and back) to generate diverse paraphrases.
  • External Commonsense Knowledge: We incorporate exemplar sentences from FrameNet that describe an object's prototypical function. For example, when processing a sentence about a "bed," the model is provided with FrameNet examples related to the "Sleep" frame to provide external context about what "use" looks like for that object.
The TOUCAN model pipeline, illustrating the integration of data augmentation and FrameNet exemplar retrieval.

Experimental Results

The TOUCAN model achieves a 72.4% accuracy and a 70.8 F1 score, representing a substantial improvement over zero-shot prompting baselines like T0++, which achieved only 46.0% accuracy. Ablation studies indicate that the inclusion of FrameNet exemplars and back-translation were the most impactful components of the model's performance, proving that external knowledge and linguistic variety are key to mastering this task.

Conclusion and Future Work

While successful, the current research is limited by a relatively small dataset of ~2,000 sentences and a focus solely on human-made artifacts. Future work will explore expanding the scope to natural objects and integrating OUC into broader narrative reasoning frameworks to see how much it improves overall story understanding.

Summary updated

BibTeX
@inproceedings{jiang2022identifying,
  author = {Tianyu Jiang and Ellen Riloff},
  title = {{Identifying Physical Object Use in Sentences}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2022},
  doi = {10.18653/v1/2022.emnlp-main.781},
}
Figure