استخراج روابط چندتايي ميان موجوديت¬ها از متون دامنه باز

In recent decades, the volume of web data has grown rapidly. A significant portion of this data exists in unstructured formats, such as web documents. With the sharp increase in unstructured data, a key question arises: how can machines effectively process an‎d utilize such data? Open relation extraction aims to automatically extract structured, machine-readable information from natural language texts in open domains (without domain-specific restrictions). Although considerable progress has been made in open relation extraction, there are still many challenges, an‎d existing systems leave much room for improvement. Most prior research has primarily focused on extracting binary relations within single sentences (i.e., relations between two entities in a sentence). However, in real-world scenarios, relations often appear in more complex forms—such as n-ary relations that span across multiple sentences. Sentence-level binary approaches tend to overlook important contextual information an‎d hinder deeper content understan‎ding. To address these challenges, this study proposes a novel system for n-ary open relation extraction at both the sentence an‎d cross-sentence (paragraph) levels. The system leverages two key techniques: Relation-aware Multi-Head Attention an‎d Relation Embedding. The combination of these components enables the system to focus on the relations extracted by SpanBERT an‎d to accurately identify the entities associated with each relation. Another important contribution of this research is the development of a new dataset, ParagOIE, which contains cross-sentence, n-ary relations from open-domain texts. This dataset is specifically designed an‎d annotated to support the precise eva‎luation of relation extraction systems at the paragraph level, an‎d it is introduced as a new benchmark in this domain. Extensive experiments conducted on stan‎dard benchmark datasets demonstrate that the proposed system outperforms current state-of-the-art systems in n-ary open relation extraction, achieving improvements of up to 10% in precision, 8% in recall, an‎d 9% in F1-score.

6