A collection of sentences is not necessarily a discourse. For a text to be meaningful, it must possess cohesion—the linguistic "glue" that connects individual statements into a unified whole. Researchers have long sought to model this glue by tracking lexical chains: sequences of semantically related words that thread through a document to maintain topical continuity.
The Semantic Fabric of Discourse
The intuition behind lexical chaining is that related concepts tend to recur. As Lance A. Miller observed in early work on business correspondence, textual cohesion is largely sustained by a limited set of mechanisms—specifically lexical repetitions, pronominal references, and collocations—which can account for over 80% of the connective links in a document.
By identifying these links, we can visualize a text not as a linear sequence of words, but as a network of overlapping threads. For example, a document discussing "hospitals" might contain a chain including "doctor," "nurse," "patient," and "clinic." Vadlapudi et al. demonstrated that these chains can be constructed automatically by linking entities across adjacent sentences using semantic relatedness derived from external knowledge bases like Wikipedia.
Measuring Textual Quality and Flow
Lexical chains provide a robust proxy for human judgments of text quality. In the high-stakes domain of automated essay scoring, Ait Khayi et al. argue that tracking these sequences allows systems to approximate how consistently a student maintains a topic. Similarly, Song et al. use identity and lexical chains to identify the "skeleton" of an argument, distinguishing between main ideas and supporting evidence based on how concepts persist across segments.
This approach extends to evaluating the output of machine translation systems. Gong et al. introduced metrics that track the repetition of content words and their stems across sentences. By comparing the overlap of these chains between a machine translation and a human reference, they can determine if the system maintains consistent terminology and linguistic flow, achieving a higher correlation with human judgment than traditional sentence-level metrics like BLEU.
Inferring Latent Structure and Meaning
Lexical cohesion also serves as a tool for uncovering hidden structures in unorganized data. Baldwin et al. showed that the recurrence of semantically related terms across forum messages can signal a direct reply relationship, allowing for the recovery of hierarchical thread structures without explicit metadata. In a similar vein, Jeong et al. used lexical cohesion to align parallel document sections, such as a story and its commentary, where shared vocabulary indicates topical links despite differences in style or length.
Beyond structure, these chains help resolve linguistic ambiguity. Li et al. found that lexical chains can distinguish between literal and idiomatic language. If the constituent words of an expression like "break the ice" participate in a strong lexical chain with the surrounding discourse (e.g., "water," "frozen," "ship"), the usage is likely literal. If the chain is absent, the usage is likely figurative.
Bridging Semantic Gaps
In specialized domains, lexical chains act as semantic shortcuts. Xiong et al. demonstrated this in scientific relation extraction, where chains link semantically related nouns across sentence boundaries. This creates a bridge for information flow, allowing models to identify relationships between entities (like a bacterium and its habitat) that are not connected by standard grammatical structures within a single sentence.
Ultimately, lexical chain analysis works because it captures the "big picture" of a conversation. As Xu et al. noted in their work on Chinese discourse, latent topic continuity provides a stronger signal for understanding how ideas relate than syntactic features alone, especially when explicit connectives are missing. By modeling these semantic threads, researchers can better approximate the human ability to perceive a unified discourse.
Key Foundational Work
Project EPISTLE: A System for the Automatic Analysis of Business Correspondence (1980) — contributed 1 concept elements to this direction.
Interactive Semantic Analysis of English Paragraphs (1969) — contributed 1 concept elements to this direction.