June 2023 Summaries

The paper by Jie Huang and Kevin Chen-Chuan Chang explores the current understanding of reasoning within Large Language Models (LLMs), highlighting techniques to improve these models' reasoning capabilities and methods to assess their performance. While LLMs have shown potential in natural language processing and exhibit reasoning abilities when scaled, it is still uncertain to what extent these models truly reason. The paper discusses various techniques such as Fully Supervised Finetuning, Prompting & In-Context Learning, and Chain of Thought, alongside suggestions for future research like hybrid methods and reasoning-enhanced training. The study emphasizes the importance of reasoning as a fundamental human trait crucial for problem-solving and decision-making, and it suggests that advancing LLMs' reasoning abilities could significantly enhance their application in complex tasks.

The paper provides a comprehensive review and comparison of methods for single-label and multi-label text classification, categorizing them into bag-of-words, sequence-based, graph-based, and hierarchical methods. It concludes that pre-trained language models consistently outperform graph-based and hierarchy-based methods, and sometimes even surpass traditional machine learning techniques like multilayer perceptrons on bag-of-words models. The study highlights the limited impact of graph-based methods, which often require more resources, and suggests that future research should benchmark against strong bag-of-words baselines and state-of-the-art pre-trained models. Additionally, it notes that simple methods such as multilayer perceptrons and logistic regression have been overlooked as substantial competitors, while sequence-based Transformers are identified as leading in text classification tasks.

The paper introduces a novel re-ranking approach for explainable recommender systems that leverages knowledge graphs to optimize explanations based on recency, popularity, and diversity, aiming to enhance explanation quality without compromising recommendation utility. Through experiments conducted on two public datasets, the study demonstrates that the proposed methods improve explanation quality while maintaining fairness across demographic groups. The authors identify three new properties that influence the perceived quality of explanations and design re-ranking strategies to optimize these properties. Additionally, the paper combines a literature review and user studies to thoroughly explore relevant explanation types, contributing by proposing new metrics for explanation quality, a suite of re-ranking approaches, and an evaluation of the impact on both recommendation utility and explanation metrics. Technologies used in the research include Python, scikit-learn, pandas, numpy, and networkx.