ArguKPE: Argument-Driven Keyphrase Extraction with Transformer-Based Semantic Alignment for Reviewer-Manuscript Matching

Keyphrase Extraction (KPE) is an important component in the comprehension of scholarly documents, as well as downstream tasks including information retrieval and reviewer-manuscript matching. The current KPE methods, such as statistical, graph-based, and new transformer-based ones, are mainly based on surface features or contextual embeddings, which fail to account for the underlying discourse structure of scientific texts. Consequently, the approaches often harvest statistically salient yet semantically inappropriate words into the main contribution of the manuscript, resulting in less-than-optimal document representation and poorer performance on downstream tasks. To overcome this shortcoming, we introduce ArguKPE, an argument-based Keyphrase Extraction model that makes an explicit effort to introduce the argumentative base of scientific texts into the extraction process. The suggested method uses argumentation mining in preference to the title and abstract to extract the important elements, i.e., the research goal and research approach, and incorporates them as structured semantic priors into the existing KPE algorithms. This definition allows the improvement of various approaches, such as statistical, graph-based, and embedding-based, by using a single argument-based scoring system. Moreover, we present ArguKPE-BERT, a transformer-based enhanced representation that integrates semantic alignment of arguments and contextual document representations. The proposed model combines semantic relevance and argumentative significance into one, achieving a more accurate and understandable Keyphrase extraction by jointly modeling both global context and discourse-level importance. Extensive testing on benchmark datasets, such as SemEval, NUS, Krapivin and Inspect, show that the proposed framework is always better than the baseline methods in all categories. In particular, when using argument-based improvements, an increase of up to 3-6% in F1-score is achieved over classical algorithms, with the proposed ArguKPE-BERT achieving a further improvement of 4-5% over powerful transformer-based baselines. These findings confirm that the use of argument structure will add complementary information on top of the traditional statistical and contextual cues. On the whole, this paper makes argument-aware modeling a powerful way to enhance Keyphrase extraction and emphasizes that it can be applied to the downstream systems as a reviewer-manuscript matching system.