Revisiting the Prepositional-Phrase Attachment Problem Using Explicit Commonsense Knowledge

Yida Xin, Henry Lieberman, Peter ChinarXiv 2021 · OpenReview

Abstract

We revisit the challenging problem of resolving prepositional-phrase (PP) attachment ambiguity. To date, proposed solutions are either rule-based, where explicit grammar rules direct how to resolve ambiguities; or statistical, where the decision is learned from a corpus of labeled examples. We argue that explicit commonsense knowledge bases can provide an essential ingredient for making good attachment decisions. We implemented a module, named Patch-Comm, that can be used by a variety of conventional parsers, to make attachment decisions. Where the commonsense KB does not provide direct answers, we fall back on a more general system that infers "out-of-knowledge-base" assertions in a manner similar to the way some NLP systems handle out-of-vocabulary words. Our results suggest that the commonsense knowledge-based approach can provide the best of both worlds, integrating rule-based and statistical techniques. As the field is increasingly coming to recognize the importance of explainability in AI, a commonsense approach can enable NLP developers to better understand the behavior of systems, and facilitate natural dialogues with end users.

Introduction

The "Prepositional-Phrase (PP) Attachment Problem" is a classic challenge in linguistics and natural language processing. It occurs when a prepositional phrase can syntactically modify multiple preceding heads, leading to ambiguity. For instance, in the sentence "John is playing the guitar with one string," does "with one string" describe the guitar (the noun) or the playing (the verb)? Humans resolve these ambiguities instantly using background knowledge, but traditional computer parsers often rely on rigid rules or opaque statistical models that lack explainability.

The PATCH-COMM Architecture

To address this, researchers developed PATCH-COMM, a hybrid system that acts as an "advisor" to conventional parsers by consulting explicit commonsense knowledge bases. The system architecture consists of three primary layers:

  1. The START Parser: A rule-based syntactic parser that identifies ambiguous PPs and pauses to request advice from PATCH-COMM.
  2. ConceptNet Integration: PATCH-COMM extracts keyword concepts and queries ConceptNet, a massive crowdsourced knowledge base. To ensure statistical stability, the system uses a "neighborhood" approach, comparing the five nearest neighbors of each word to calculate robust relationship weights.
  3. RetroGAN-DRD (The Fallback): To solve the "insufficiency problem" of knowledge bases, the system uses a neural-symbolic fallback. This component utilizes fastText embeddings and a CycleGAN-inspired architecture to infer likely relations between concepts not explicitly present in the KB.
ConceptNet query results showing the 'HasA' relation between string and guitar used to resolve attachment.

Experimental Evaluation

The system was evaluated using a subset of the Belinkov et al. (2014) English dataset, specifically focusing on the preposition "with." The test set comprised 48 data points with up to three candidate heads.

Key Results

  • Baseline Accuracy: 39.2% (calculated by random assignment).
  • PATCH-COMM Accuracy: 45.8%.
  • Near-Misses: In 27.1% of incorrect cases, the correct answer's weight was within 0.1 of the chosen answer, indicating the system was often close to the correct reasoning.

While the accuracy is lower than state-of-the-art black-box models (which reach ~88%), the authors emphasize that PATCH-COMM is designed for transparency and integration into larger "story-understanding" systems like Genesis.

Explainability and Future Work

A primary goal of this research is explainability. Unlike pure neural models, PATCH-COMM provides human-readable justifications for its decisions (e.g., "I attached 'string' to 'guitar' because the knowledge base says a guitar has strings").

The resulting semantic triplets showing the correct attachment of the prepositional phrase to the noun 'guitar'.

By producing explicit triplets, the system allows for better debugging and user interaction. Moving forward, the researchers plan to extend this approach to co-reference resolution and broader Discourse Understanding, where global story context (e.g., knowing a character is a toddler) can override local linguistic defaults to resolve complex narrative puzzles.

Summary updated

Abstract

We revisit the challenging problem of resolving prepositional-phrase (PP) attachment ambiguity. To date, proposed solutions are either rule-based, where explicit grammar rules direct how to resolve ambiguities; or statistical, where the decision is learned from a corpus of labeled examples. We argue that explicit commonsense knowledge bases can provide an essential ingredient for making good attachment decisions. We implemented a module, named Patch-Comm, that can be used by a variety of conventional parsers, to make attachment decisions. Where the commonsense KB does not provide direct answers, we fall back on a more general system that infers "out-of-knowledge-base" assertions in a manner similar to the way some NLP systems handle out-of-vocabulary words. Our results suggest that the commonsense knowledge-based approach can provide the best of both worlds, integrating rule-based and statistical techniques. As the field is increasingly coming to recognize the importance of explainability in AI, a commonsense approach can enable NLP developers to better understand the behavior of systems, and facilitate natural dialogues with end users.
BibTeX
@misc{xin2021revisiting,
  author = {Yida Xin and Henry Lieberman and Peter Chin},
  title = {{Revisiting the Prepositional-Phrase Attachment Problem Using Explicit Commonsense Knowledge}},
  year = {2021},
  eprint = {2102.00924},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2102.00924},
}
Figure