Cross-Sentence Relation

Unlike intra-sentence extraction, cross-sentence relations demand that models aggregate and retain contextual information across discourse units. This requires mechanisms to track entity mentions through coreference chains and understand logical connectors (e.g., 'however', 'therefore') that signal relationships between separate clauses. Models must build a document-level representation that captures the global narrative flow, not just local syntactic patterns.

Entities in a cross-sentence relation can be separated by dozens or hundreds of tokens, creating a long-range dependency problem. Architectures must overcome the limitations of fixed context windows. Techniques include:

• Transformer memory mechanisms that cache past key-value pairs
• Graph-based document models that connect entity mentions as nodes with edges representing sentence adjacency or coreference
• Hierarchical attention networks that first encode sentences, then model inter-sentence interactions

Cross-sentence relation extraction is heavily dependent on accurate coreference resolution. When an entity is introduced in one sentence ('Acme Corp. announced a new product') and referred to with a pronoun in another ('It will launch next quarter'), the system must resolve 'It' to 'Acme Corp.' before the temporal relation can be extracted. Errors in coreference linking cascade directly into relation extraction failures.

Advanced cross-sentence systems leverage Rhetorical Structure Theory (RST) to parse how clauses and sentences relate to one another. By identifying discourse relations such as Cause-Effect, Condition, Elaboration, or Contrast between text spans, models gain a structural scaffold for predicting semantic relationships. A 'Contrast' discourse relation between two sentences, for example, strongly suggests an opposing relationship between the entities they contain.

A common architectural pattern involves constructing a heterogeneous document graph where nodes represent entities, mentions, and sentences, while edges encode syntactic dependencies, coreference links, and sentence adjacency. Graph Neural Networks (GNNs) then propagate information across this structure, allowing a relation classifier to consider evidence aggregated from multiple sentences. This approach explicitly models the multi-hop reasoning path required for cross-sentence extraction.

Cross-sentence relations often require implicit reasoning beyond explicit textual cues. For example, 'The CEO resigned. The stock price plummeted.' implies a causal relation, but the text never states it directly. Models must incorporate commonsense knowledge and perform abductive inference to bridge these logical gaps. This is a key differentiator from intra-sentence extraction, where syntactic proximity often makes the relationship surface-level.