Dependency Paths

The minimal syntactic route connecting two entities in a dependency tree. The SDP filters out irrelevant words and focuses on the grammatical core of the relationship.

• Path Length: Typically 1-5 edges for related entities
• Example: In 'The CEO of Apple announced a product', the SDP between 'CEO' and 'Apple' is CEO <-nmod- of -case-> Apple
• Key Insight: The sequence of dependency relations along the SDP is a strong feature for classifying the semantic relation type

A variant of the SDP that includes the actual words (lemmas) at each node along the path, not just the dependency labels. This captures both syntactic structure and lexical semantics.

• Components: Word lemmas, POS tags, and dependency labels combined into a single path representation
• Example: CEO/noun<-nmod/Apple/noun encodes both the words and their relationship
• Use Case: Kernel methods and neural models use these paths as input features for relation classification

A kernel function that computes the similarity between two entity pairs by comparing their dependency path structures. This allows SVMs and other kernel methods to operate directly on syntactic structures.

• Subsequence Kernels: Count common subsequences of dependency relations between paths
• Convolution Kernels: Decompose the parse tree into substructures and compare them recursively
• Advantage: Captures syntactic generalizations without manual feature engineering

The deepest node in the dependency tree that is an ancestor of both entities. The LCA often represents the syntactic head that governs the relationship between the two entities.

• Path Construction: The SDP is formed by concatenating the path from entity1 to the LCA and from the LCA to entity2
• Typical LCAs: Verbs (for event relations), prepositions (for spatial relations), or conjunctions
• Example: In 'Microsoft acquired LinkedIn', the verb 'acquired' is the LCA connecting the subject and object

The process of abstracting dependency paths into general patterns by replacing specific words with their POS tags or entity types. This enables the model to recognize the same relationship expressed with different vocabulary.

• Entity Masking: Replace 'Microsoft' and 'LinkedIn' with [ORG] tags
• Pattern Example: [ORG] <-nsubj- acquired -dobj-> [ORG] generalizes to any acquisition relation
• Benefit: Dramatically reduces sparsity and improves recall on unseen entity pairs

Syntactic paths that pass through negation markers, which fundamentally alter the semantic relationship. Detecting these paths is critical for distinguishing affirmative from negated relations.

• Negation Markers: Words like 'not', 'never', 'no', and their dependency modifiers
• Path Impact: A path through a neg dependency edge inverts the truth value of the relation
• Example: 'Apple did not acquire Beats' contains a negation edge that nullifies the acquisition relation

The foundational NLP task that produces the syntactic tree structure upon which dependency paths are defined. A dependency parser analyzes the grammatical structure of a sentence, establishing directed binary relations between a head word and its dependent words. The resulting tree is a graph where nodes are words and edges are typed grammatical relations (e.g., nsubj, dobj, prep). The shortest path through this graph connecting two entity mentions is the dependency path.

The primary downstream application for dependency paths. RE is the task of automatically identifying and classifying semantic relationships between named entities in text. Dependency paths serve as a powerful, interpretable feature for RE classifiers. For example, the path between a PERSON and an ORGANIZATION in 'Alice, CEO of Acme Corp' follows the pattern PERSON → appos → CEO → prep → of → pobj → ORGANIZATION, which strongly indicates an employed_by or leadership relation.

The specific syntactic route connecting two entities in a dependency tree, defined as the minimal set of edges between them. The SDP filters out irrelevant clauses and focuses the model on the core grammatical link. Key properties include:

• Directionality: The path captures the order of traversal from the source entity to the target entity.
• Edge Types: The sequence of grammatical relations (e.g., nsubj → dobj) forms a feature vector.
• Lexicalization: The words along the path can be included as features, forming a lexicalized dependency path.

Modern neural architectures that model the entire dependency tree as a graph, moving beyond simple path features. Instead of extracting a single SDP, a GNN learns to aggregate information from a node's local neighborhood across the parse tree. This allows the model to capture inter-entity dependencies and long-range syntactic context that a single path might miss, significantly improving performance on document-level and cross-sentence relation extraction.

Classic machine learning approaches that compute the similarity between two syntactic structures directly. A dependency tree kernel implicitly maps a sentence's parse tree into a high-dimensional feature space, counting common substructures like subtrees or subset trees. This allows an SVM classifier to leverage the full syntactic context, including the SDP, without manual feature engineering, forming the historical bridge between rule-based and neural RE.

A semantic relationship where the two entities appear in different sentences within the same document. Standard dependency paths fail here because there is no single syntactic tree spanning multiple sentences. Solving this requires document-level discourse parsing or coreference resolution to create a virtual link, or the use of transformer-based models that can attend across sentence boundaries without explicit syntactic structure.