Abstract
Complex questions in reading comprehension tasks require integrating information from multiple sentences. In this work, to answer such questions involving temporal and causal relations, we generate event graphs from text based on dependencies, and rank answers by aligning event graphs. In particular, the alignments are constrained by graph-based reasoning to ensure temporal and causal agreement. Our focused approach self-adaptively complements existing solutions; it is automatically triggered only when applicable. Experiments on RACE and MCTest show that state-of-the-art methods are notably improved by using our approach as an add-on.
Introduction
While modern neural models excel at extracting facts from single sentences, they often struggle with questions requiring multi-sentence reasoning, particularly those involving chronological order and cause-and-effect. For instance, if a passage states "Susan hung up balloons before her friends arrived," a model must understand the temporal sequence to correctly answer what happened first. Most existing models either ignore these semantic links or use sliding windows that fail to capture the underlying logical structure. This paper introduces GTCR (Graph-based Temporal-Causal Reasoning) to bridge this gap by explicitly modeling event relationships across a passage.
The GTCR Framework
The GTCR system treats reading comprehension as a graph-matching problem. The architecture follows a structured pipeline that transforms raw text into symbolic representations suitable for logical reasoning.
The process consists of four primary stages:
- Event Graph Generation: Extracting "events" (predicates and arguments) using dependency parsing and identifying temporal (before, after) and causal (result, cause) relations.
- QA Graph Generation: Merging the question and candidate answer into a compact "QA graph."
- Graph Alignment: Using 0-1 Linear Programming to find the optimal match between the QA graph and the passage graph.
- Graph-based Reasoning: Enforcing logical consistency. If the passage graph indicates Event A preceded Event B, but the candidate answer implies the reverse, the system identifies a conflict and penalizes that alignment.
Event Graph Representation and Reasoning
The core innovation lies in representing stories as global maps of events. By looking for syntactic patterns within sentences and connective adverbs (e.g., "therefore," "later") between sentences, GTCR builds a comprehensive temporal-causal web.
This symbolic approach allows the model to move beyond simple word similarity. By enforcing temporal and causal "agreement," the system can filter out answers that are linguistically plausible but logically impossible given the story's timeline. Furthermore, GTCR is designed as a self-adaptive add-on; it only triggers when it detects a temporal/causal question and when the quality of the dependency parse provides high confidence.
Experimental Results
The researchers evaluated GTCR on two major benchmarks: RACE-M (English exam questions for middle school students) and MCTest (fictional stories requiring multi-sentence reasoning).
Key findings include:
- Performance Boost: On questions specifically involving time and cause, GTCR outperformed state-of-the-art neural models, such as the Stanford Attentive Reader, by 4–10%.
- Synergy with Neural Models: When used as an add-on, it consistently improved overall accuracy. For example, it boosted the Gated-Attention (GA) Reader on RACE-M from 43.73% to 44.64%.
- Ablation Study: Removing the reasoning component (the logic checks) led to significant accuracy drops (up to 13% on some datasets), confirming that the graph-based logic is the primary driver of performance.
Conclusion and Future Work
GTCR demonstrates that symbolic structures like graphs and logic can provide explainability and precision that pure neural networks often lack in complex reasoning tasks. However, the system remains sensitive to the quality of initial dependency parses. Future work will focus on handling implicit relations (where logic is implied but not stated) and moving toward "one-to-many" matching strategies to better handle sentences that summarize multiple events.
Summary updated