Inferring Temporally-Anchored Spatial Knowledge from Semantic Roles

Eduardo Blanco, Alakananda VempalaNAACL 2015

TL;DR

This paper introduces a novel framework to infer spatial relationships (where things are and aren't) from existing semantic role annotations in text. By leveraging PropBank-style semantic roles, the system can automatically extract and temporally anchor spatial knowledge, revealing that such inferences are common and intuitive for humans.

What Problem Does This Paper Address?

Understanding natural language often requires more than just identifying the core relationships between words (like who did what to whom). Specifically, understanding where entities are located and when these locations hold true is crucial for a deeper comprehension of text. While semantic role labeling (SRL) has made significant strides in identifying relationships between verbs and their arguments, it often falls short of capturing nuanced spatial information and its temporal dynamics. For instance, knowing "John was incarcerated at Shawshank prison" tells us about John's location during incarceration, but it doesn't explicitly state that he wasn't there before or after. This paper aims to bridge this gap by inferring this "additional spatial knowledge" and anchoring it in time relative to the event described by the verb.

Methods

The core of the proposed method involves a two-step process:

  1. Deterministic Generation of Potential Spatial Knowledge: An algorithm (Algorithm 1) systematically iterates through sentences, identifying verbs with location arguments (ARGM-LOC) and other numbered arguments (ARG0, ARG1, etc.). It then generates potential spatial inferences about whether these arguments are located at the ARGM-LOC before, during, or after the verb's event. A crucial step here is a is_valid(x, y) check to avoid nonsensical inferences where one entity is contained within another.
  2. Manual Annotation and Model Training: To validate these inferences, the authors conducted a manual annotation effort. Annotators were presented with sentences and potential spatial questions, responding with degrees of certainty (e.g., "certainly yes," "probably yes," "unknown"). This annotation process was refined based on initial feedback to capture nuanced human understanding. Finally, a supervised machine learning model (SVM with RBF kernel) was trained using a rich set of features derived from lexical, syntactic, and semantic information to predict these temporally-anchored spatial relationships.

Datasets

The study utilizes the OntoNotes corpus, a large, multi-genre corpus that includes PropBank-style semantic role annotations. For their annotation effort, the authors selected 200 sentences from OntoNotes that contained at least one ARGM-LOC role. This subset was used to generate potential spatial knowledge and then manually annotated. The annotated data was split into 80% for training and 20% for testing.

Key Insights and Ideas

  • Ubiquity of Spatial Inferences: The annotation effort revealed that inferring spatial knowledge from semantic roles is not an edge case but a common and intuitive process for humans. Annotators were able to provide answers with high confidence for a significant portion of the generated questions.
  • Temporal Anchoring is Key: Simply knowing a location isn't enough; understanding when that location holds true (before, during, or after an event) is crucial for a complete understanding. The framework explicitly addresses this temporal dimension.
  • Beyond Motion Verbs: While it's intuitive to infer spatial information from motion verbs (e.g., "went to Paris"), the paper demonstrates that verbs like "incarcerated" also yield significant spatial inferences. This broadens the scope of where spatial knowledge can be extracted.
  • Uncertainty is Natural: Humans often express spatial knowledge with degrees of certainty. The annotation scheme, allowing for "probable" answers, reflects this reality and was crucial for achieving high inter-annotator agreement.

Key Results

  • The system infers over 44% of potential spatial knowledge with certainty (certYES and certNO labels).
  • Annotators were able to infer spatial knowledge with some degree of certainty in over 94% of cases where an ARGM-LOC role was present.
  • The best performing model, combining lexical, syntactic, and semantic features, achieved a weighted F-measure of 0.55.
  • Inferences for the "during" temporal anchor were significantly higher (0.65 F-measure) compared to "before" (0.41) and "after" (0.45) anchors, suggesting that spatial relationships are often most strongly tied to the event itself.

Limitations and Future Work

  • Type (2) Additional Spatial Knowledge: The current framework focuses on inferring spatial knowledge between arguments of verbs (type 1). The authors acknowledge that inferring spatial knowledge where entities are not direct semantic roles of a verb (type 2) is a direction for future work.
  • Random Splitting of Instances: The authors note that randomly splitting instances from the same sentence between train and test sets would be unfair, as sentence-level context is important. They thank an anonymous reviewer for pointing this out.
  • Scope of Features: While the feature set is rich, further exploration of more sophisticated semantic features could potentially improve performance.
  • Generalizability: While the framework is designed to be pluggable into existing NLP pipelines, further evaluation on diverse domains and genres would be beneficial.

Summary updated

BibTeX
@inproceedings{blanco2015inferring,
  author = {Eduardo Blanco and Alakananda Vempala},
  title = {{Inferring Temporally-Anchored Spatial Knowledge from Semantic Roles}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2015},
}

Loading related content...

Figure