Introduction
Math Word Problems (MWPs) represent a significant challenge in Natural Language Processing (NLP), requiring a functional understanding of narrative context rather than simple keyword matching. Traditional approaches often fall into two categories: brittle rule-based systems that struggle with linguistic ambiguity, and "black-box" statistical models that are sensitive to irrelevant "noise" and lack transparency. This paper, presented at IJCAI-16, introduces a hybrid system designed to bridge this gap. By utilizing a tag-based logic framework, the solver calculates answers to English MWPs while filtering out irrelevant information and providing human-readable explanations.
System Architecture
The proposed system transforms raw text into a structured "Reasoning Chain" through a five-module pipeline.
- Language Analyzer: Employs the Stanford CoreNLP suite for dependency parsing and co-reference resolution.
- Solution Type Classifier (STC): Uses a linear SVM to predict the required mathematical operation (e.g., addition, subtraction) based on linguistic features.
- Logic Form Converter (LFC): Converts text into a variant of First-Order Logic (FOL), attaching semantic "tags" to numerical quantities.
- Inference Engine (IE): Executes the mathematical operations, using tags as filters to distinguish between relevant and irrelevant data.
- Explanation Generator: Produces human-comprehensible step-by-step explanations using the reasoning chain and pre-defined templates.
The Tag-Based Logic Framework
The core innovation of this system is the Logic Form Converter. Unlike standard solvers that merely extract scalars, this system preserves the "physical meaning" of numbers by associating them with agents (who), verbs (action), and objects (what).
For example, in a problem where Fred picks 36 limes and Nancy drops 18, the system generates logic forms such as quan(q1, #, lime)=36 & verb(q1, pick) & nsubj(q1, Fred). When the question asks for the "total picked," the Inference Engine uses these tags to select only the quantities associated with the "pick" verb, effectively ignoring the "dropped" quantity.
Key Insights and Results
The system's ability to handle "distractor" information is a major advancement over purely statistical solvers. By preserving metadata through tags, the system mimics human-like situational awareness. Furthermore, because the solution path is based on logic forms rather than opaque neural weights, the reasoning process is inherently transparent. The system can generate reasoning trees and textual explanations, making it a viable candidate for educational tools like computer math tutors.
Limitations and Future Outlook
While effective, the system is heavily dependent on upstream NLP accuracy; errors in dependency parsing or solution classification can propagate through the pipeline. Additionally, representing highly abstract mathematical concepts in First-Order Logic remains computationally challenging. Future work aims to expand this tag-based approach to more complex problem types and integrate broader domain knowledge.
Summary updated