In the late 1980s, natural language processing (NLP) faced a significant hurdle in discourse analysis: understanding how individual sentences coalesce into a coherent argument. Robin Cohen’s 1987 paper, "Interpreting Clues in Conjunction with Processing Restrictions in Arguments and Discourse," remains a seminal work for its formalization of "clue words"—connectives like however, as a result, and in addition—as structural signals that guide computational parsing.
The Computational Challenge of Discourse
The primary goal of discourse analysis in this context is to construct a "representation tree." In such a tree, nodes represent individual propositions, and edges represent "evidence" relationships, where one proposition supports another.
Without structural hints, a computer must exhaustively test every new sentence against all "eligible" prior sentences to determine its logical placement. This process relies on an "Evidence Oracle"—a conceptual module that judges logical support. Because calls to this oracle are computationally expensive, an unguided search is inefficient and prone to errors, particularly when speakers use non-linear rhetorical strategies.
The Clue Taxonomy
Cohen’s breakthrough was categorizing linguistic connectives based on the specific structural relationships they imply between a new proposition ($S$) and a prior one ($P$). By mapping these "clue words" to tree operations, the system can bypass the exhaustive search of the Evidence Oracle.
The taxonomy identifies four primary categories:
- Parallel (e.g., "in addition"): Indicates $S$ is a sibling ("brother") to $P$.
- Detail (e.g., "in particular"): Indicates $S$ is evidence ("son") for $P$.
- Inference (e.g., "as a result"): Indicates $S$ is the claim ("father") supported by $P$.
- Summary (e.g., "in sum"): Indicates $S$ is the "father" to multiple preceding "sons."
Pruning the Search Space
By integrating this taxonomy into a tree-building algorithm, the system uses clues to prune the search space. For example, if a "parallel" clue is detected, the algorithm immediately rules out any nodes that are not siblings of the current focus. This significantly reduces the number of calls to the Evidence Oracle.
Furthermore, this approach allows for early detection of logical incoherence. If a speaker uses a summary clue like "In sum..." without having provided any preceding evidence, the algorithm can trigger an INTERRUPT-DISCOURSE signal immediately, rather than waiting until the end of the analysis.
Handling Non-Linearity and Re-attachment
Standard discourse parsers often struggle with "exceptional transmissions," such as when a speaker returns to a topic mentioned several sentences prior. Clues provide the necessary context to navigate these shifts, allowing the algorithm to reopen "closed" parts of the argument tree.
Cohen also highlights the efficiency of re-attachment strategies. It is computationally "cheaper" to test if a new proposition is a son of the current node before testing if existing nodes should be re-attached as sons of the new proposition. This logic is based on the transitive nature of evidence: if $A$ supports $B$, then any evidence supporting $A$ likely supports $B$ as well.
Conclusion and Legacy
While the paper relies on an idealized "Evidence Oracle"—a challenge that modern NLP now addresses with large language models and entailment systems—its formalization of discourse structure remains highly influential. By mapping linguistic categories to computational tree operations, Cohen demonstrated how structural "hints" can transform an intractable search problem into an efficient, robust parsing process. Future work suggested extending these concepts to non-verbal clues, such as intonation in spoken discourse.
Summary updated