When a parser processes a sentence like "The engineer inspected the pump with the telescope," it faces a structural crossroads. Does the telescope belong to the pump, or is it the tool used for the inspection? Standard formal grammars merely enumerate these valid structural pathways, leaving a parser paralyzed by choice.
To resolve such local ambiguities efficiently, both human minds and computational parsers rely on structural parsing preferences—algorithmic tie-breakers that favor certain tree shapes over others. Crucially, these principles operate primarily on the geometry of the syntax tree itself, allowing the parser to make rapid, reliable decisions before invoking expensive semantic analysis.
The Locality Principle: Right Association and Low Attachment
Why do we naturally prefer certain attachments over others? The first core intuition is locality: incoming words prefer to attach to the closest available structure currently open in working memory. This is known variously as Right Association, Low Attachment, or Structural Recency.
In Experience with an Easily Computed Metric for Ranking Alternative Parses, George E. Heidorn operationalized this by designing a metric that mathematically favors deeper, right-branching structures. By calculating tree depth and forcing new constituents to attach to the lowest possible node in the existing parse tree, this metric successfully mirrors human grammatical intuition in the vast majority of structural ties.
Shallow Attachment (Less Preferred) Deep/Low Attachment (Preferred)
VP VP
/ \ / \
VB NP VB NP
(saw) / \ (saw) / \
NP PP NP NP
(the man) (with the telescope) (the man) / \
N PP
(pump) (with the telescope)
This preference for proximity minimizes cognitive load. The "Attach Low and Parallel" principle articulated by the authors of Two Principles of Parse Preference argues that the parser prioritizes the nearest available attachment site to keep local dependencies tightly bound. When evaluating where to place post-modifier prepositional phrases, Greg Whittemore et al. demonstrated in Empirical Study of Predictive Powers od Simple Attachment Schemes for Post-Modifier Prepositional Phrases that this structural recency heuristic drastically reduces tree complexity, serving as a highly predictive baseline for human behavior.
The Economy Principle: Minimal Attachment and Simplicity
If the locality principle is about staying close, the economy principle is about staying simple. When confronted with an ambiguous phrase, the parser defaults to the structure that requires the fewest structural nodes or subtrees—a concept known as Minimal Attachment.
In An Analysis on Chinese EFL Learners’ Understanding of “Direct Object/Subject” Garden Path Sentences, Yue Bao explores how this bias can lead readers into "garden-path" traps. The language faculty defaults to the simplest possible grammatical structure, minimizing the number of nodes in the syntactic tree. For instance, in a sentence like "While the man was hunting the deer ran into the woods," the parser immediately swallows "the deer" as the direct object of "hunting" because building a new clausal boundary would require introducing additional structural nodes. Only when encountering "ran" is the reader forced to mentally re-parse.
This node-minimization strategy extends beyond standard constituent trees to more complex grammar formalisms:
- Derivation Trees: In Parsing preferences with Lexicalized Trey Adjoining Grammars exploiting the derivation tree, Alexandra Kinyon shifted the focus from the final surface tree to the derivation tree (the history of how the sentence is constructed). Kinyon demonstrated that a simplicity principle preferring derivation trees with the minimum number of nodes naturally prioritizes idiomatic readings over literal ones, because multi-word idioms are packaged as single complex nodes rather than fragmented individual components.
- Unsupervised Subtrees: Even in data-driven, unsupervised frameworks, simplicity remains a dominant force. As shown by Rens Bod in Is the End of Supervised Parsing in Sight?, an effective way to parse without a grammar is to select the derivation that requires the fewest number of previously seen subtrees. This approach naturally favors larger, more frequent structural fragments over many small, fragmented rules.
Mechanics of the Structural Tie-Breaker
How do these abstract geometric preferences manifest in actual parsing mechanics?
Consider a standard shift-reduce parser, which must constantly decide whether to pull a new token onto its stack (shift) or combine existing stack elements into a higher-level phrase (reduce). In Sentence Disambiguation by a Shift-Reduce Parsing Technique, Stuart M. Shieber proved that complex semantic rules are not required to achieve human-like disambiguation. By simply hardcoding a preference to shift rather than reduce, the parser naturally delays closing phrases. This shift-priority heuristic automatically forces incoming words to attach to the most recently processed, lowest constituent on the stack—elegantly emergentizing both Right Association and Minimal Attachment from a simple control loop.
Input: [the, pump, with, the, telescope]
Stack: [inspected] [the pump] <- Parser conflict: Reduce 'inspected the pump'
OR Shift 'with'?
Decision: SHIFT 'with' <- Delaying reduction forces the incoming PP
to attach low to 'the pump'.
When these initial structural gambles fail, the parser must perform a reanalysis. Patrick Sturt modeled this incrementally in Incorporating “Unconscious Reanalyses” into an Incremental, Monotonic Parser, utilizing a bottom-up search strategy that favors core argument attachments over optional adjuncts, implementing a "tree-lowering" operation to resolve minor garden paths without crashing the parse state.
Ultimately, these disparate structural heuristics can be unified under a single computational architecture. The authors of A Parser based on Connectionist Model proposed that parsing preferences are an emergent property of activation levels within a neural network. Instead of rigid symbolic rules, competing structural interpretations accumulate activation based on frequency and recency. The parse tree with the highest activation is inherently preferred, providing a single, fluid mechanism where both minimal attachment and right association naturally coexist.
Key Foundational Work
Sentence Disambiguation by a Shift-Reduce Parsing Technique (1983) — contributed 3 concept elements to this direction.
Experience with an Easily Computed Metric for Ranking Alternative Parses (1982) — contributed 2 concept elements to this direction.