This AI Paper by the University of Michigan Introduces MIDGARD: Advancing AI Reasoning with Minimum Description Length

Structured commonsense reasoning in natural language processing involves automated generating and manipulating reasoning graphs from textual inputs. This domain focuses on enabling machines to understand and reason about everyday situations as humans would, translating natural language into interconnected concepts that mirror human logical processes.

One of the fundamental challenges in this field is the difficulty of accurately modeling and automating commonsense reasoning. Traditional methods often need help with error propagation and robust mechanisms for correcting inaccuracies during graph generation, which can result in incomplete or incorrect reasoning structures. Improving methods is critical to enhance the accuracy and reliability of automated reasoning systems.

Existing research in structured commonsense reasoning includes frameworks like COCOGEN, which employs programming scripts as prompts to guide LLMs in generating structured outputs. Despite improvements, COCOGEN still needs help with style mismatch and error propagation. The self-consistency framework enhances model reliability by aggregating common results from multiple samples. Furthermore, training-based methods utilize verifiers and re-rankers to refine sample selection, aiming to align outputs more closely with human judgment. These methods demonstrate the evolving strategies to tackle the inherent complexities of reasoning in natural language processing.

Researchers from the University of Michigan have introduced MIDGARD, a novel framework utilizing the Minimum Description Length (MDL) principle to enhance structured commonsense reasoning. Unlike previous methods that rely on single sample outputs, which may propagate errors, MIDGARD synthesizes multiple reasoning graphs to produce a more accurate and consistent composite graph. This unique approach minimizes error propagation typical in autoregressive models. It ensures the precision of the resultant reasoning structure by focusing on the recurrence and consistency of graph elements across samples.

MIDGARD’s methodology involves generating diverse reasoning graphs from natural language inputs using a Large Language Model like GPT-3.5. These graphs are then processed to identify and retain commonly occurring nodes and edges, discarding outliers using the MDL principle. The consistency and frequency of these elements are rigorously analyzed to ensure they reflect correct reasoning patterns. Datasets used in benchmarking MIDGARD include argument structure extraction and semantic graph generation tasks, which significantly outperform existing models by demonstrating enhanced accuracy and robustness in reasoning graph construction.

MIDGARD demonstrated significant improvements in structured reasoning tasks. In the argument structure extraction task, MIDGARD increased the edge F1-score from 66.7% to 85.7%, indicating a significant reduction in error rates compared to baseline models. Moreover, MIDGARD consistently achieved higher accuracy in semantic graph generation, with performance gains reflected across various benchmarks. These quantitative results validate MIDGARD’s effectiveness in synthesizing more accurate and reliable reasoning graphs from multiple samples, showcasing its advancement over traditional single-sample-based approaches in natural language processing.

To conclude, the MIDGARD framework represents a significant advancement in structured commonsense reasoning by employing the Minimum Description Length principle to aggregate multiple reasoning graphs from large language models. This approach effectively reduces error propagation and improves the accuracy of reasoning structures. MIDGARD’s robust performance across various benchmarks demonstrates its potential to enhance natural language processing applications. It is a valuable tool for developing more reliable and sophisticated AI systems capable of understanding and processing human-like logical reasoning.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our Telegram Channel, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our 42k+ ML SubReddit