Towards Understanding Natural Language: Semantic Parsing, Commonsense Knowledge Acquisition and Applications

Arpit SharmaIJCAI 2016

Introduction

While modern Artificial Intelligence has achieved remarkable success in pattern matching and statistical prediction, true "Natural Language Understanding" (NLU) remains an elusive goal. Most current systems struggle with the deeper logic of language, often failing when a sentence requires world knowledge or an understanding of causal relationships. This research addresses two primary hurdles in the field: the lack of semantic depth in standard parsing and the "commonsense gap"—the background knowledge humans use to resolve ambiguity.

This paper introduces a framework that combines K-Parser, a semantic parser that maps text to rich graphical representations, with a novel method for automatically extracting Event-based Conditional Commonsense (ECC). By integrating these tools with declarative logic programming, the research demonstrates a viable path toward solving complex linguistic puzzles like the Winograd Schema Challenge (WSC).

The K-Parser Framework

At the core of this approach is K-Parser, a system designed to move beyond simple syntactic analysis. Unlike standard parsers that focus on grammatical structure, K-Parser translates English sentences into a Directed Acyclic Graph (DAG).

In this representation:

  • Nodes represent words and their conceptual classes (e.g., identifying "John" as a member of the class "person").
  • Edges represent semantic relations between these entities.
  • Event Mentions are rooted at verbs, capturing not just who did what, but how different events relate to one another through causality and temporal sequence.

One of the primary advantages of K-Parser over existing standards like Abstract Meaning Representation (AMR) is its attention to detail. While AMR often ignores tenses and entity quantification, K-Parser preserves these nuances, which are frequently essential for logical reasoning and determining the specific roles entities play across multiple sentences.

Bridging the Commonsense Gap

To understand language, a system must know more than just the words on the page; it must understand the "rules" of the world. This research introduces Event-based Conditional Commonsense (ECC), a method for automatically extracting "if-then" logic from large text repositories.

For example, consider the sentence: "John was bullying Tom so we rescued him." A human knows "him" refers to Tom because victims—not bullies—are typically rescued. The ECC algorithm allows the system to learn this rule automatically: IF A bullying B causes T to rescue Z, THEN Z is likely B. By extracting these rules from massive datasets, the system reduces the need for manual "knowledge engineering," where humans must hand-code every possible world rule.

Reasoning with Answer Set Programming (ASP)

To process the rich data provided by K-Parser and the rules extracted via ECC, the framework utilizes Answer Set Programming (ASP). ASP is a form of declarative logic programming that allows the system to reason through complex constraints. When faced with an ambiguous pronoun in a sentence, the system uses ASP to combine the specific semantic graph of that sentence with its library of ECC rules to find the most logically consistent interpretation.

Evaluation and Applications

The framework was primarily benchmarked using the Winograd Schema Challenge (WSC). The WSC consists of sentence pairs that differ by only one or two words but require commonsense reasoning to resolve pronoun references. The system successfully solved a subset of these challenges, proving that the combination of deep semantic parsing and automated commonsense acquisition can outperform purely statistical models in high-logic scenarios.

Beyond theoretical benchmarks, the research explored practical applications in the medical field. An early version of the parser was utilized to analyze the language patterns of patients to help determine if they were suffering from Alzheimer’s disease, demonstrating the framework's versatility across different domains.

Limitations and Future Directions

Despite its successes, the system faces several ongoing challenges:

  1. Preposition Disambiguation: Understanding the exact sense of prepositions (e.g., the different meanings of "with" or "for" in varying contexts) remains a hurdle for the underlying parser.
  2. Standardization: To facilitate better benchmarking against state-of-the-art models, future work involves translating K-Parser’s unique DAG output into the standard AMR format.
  3. Actionable Knowledge: The research team is currently developing an "Action Language Framework" to transition from static commonsense rules to a more dynamic reasoning system, making the extracted knowledge even more effective for real-time NLU applications.

Conclusion

By focusing on causality and the integration of background world logic, this research provides a robust framework for moving AI from simple text processing toward genuine understanding. The synergy between K-Parser’s semantic depth and ECC’s automated knowledge acquisition offers a promising scalable solution to the most persistent problems in natural language processing.

Summary updated

BibTeX
@inproceedings{sharma2016towards,
  author = {Arpit Sharma},
  title = {{Towards Understanding Natural Language: Semantic Parsing, Commonsense Knowledge Acquisition and Applications}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure