The comparative question identification task involves detecting and classifying questions that aim to compare entities, such as products or services. This task is important for applications like review summarization, recommendation systems, and opinion mining. Researchers have developed methods involving natural language processing techniques to accurately identify and categorize comparative questions from user-generated content. These methods often require syntactic and semantic analysis to understand the nuances and patterns in language that indicate comparisons.

Recent advancements in this field include the use of machine learning models and neural networks to improve the accuracy of identifying comparative questions. These models are trained on large datasets to recognize both explicit and implicit comparisons. Challenges in this area include handling context-dependent comparisons and refining algorithms to better capture the intent behind complex linguistic structures. This work enhances applications by providing deeper insights into user preferences and opinions through effective comparison detection.

Object and Aspect Labeling is a sequence tagging task aimed at identifying entities such as objects, aspects, and predicates in comparative questions.

There are at least three different datasets for this task. The dataset by Li et al. focuses on extracting objects of comparison but is not publicly available. The other two datasets, by Beloucif et al. and Bondarenko et al., have different annotation schemas and tags. Another dataset by Chekalina et al. includes labeled affirmative sentences.

Stance classification is one of the crucial subtasks of the whole pipeline, as we select relevant arguments and detect, in favor of which object the choice is made thanks to the class assigned at this step.

There is a dataset published by Panchenko et al. along with the baseline classifier, which is outperformed by Ma et al. Bondarenko et al. also introduces datasets and tests them on several models. However, their text excerpts are quite large and are not consistent with the argumentative sentences for our task.

Comparative Summary Generation is a relatively recent and less widespread task.

To the best of our knowledge, only two papers introduce datasets and baseline approaches. Chekalina et al. (2021) presents comparative questions with their best answers from Yahoo! Answers and tests several unsupervised approaches, including CTRL and template-based answers. Yu et al. (2023) pre-trains LLMs for comparative reasoning using prompts and introduces a dataset called “Diffen” that works with summarizing arguments for an object pair, though it is not publicly available.