Does the algorithm require neural-network training?

The historical tree-search method does not, though a surrounding pipeline may use learned parsing or other components.

Is the first compatible candidate always correct?

No. Structural preference and compatibility do not fully resolve meaning.

Can any parse representation be substituted unchanged?

No. The procedure assumes particular constituency structures; adaptations need explicit mapping and testing.

AI Glossary / Natural language and speech / Intermediate

Hobbs algorithm

Definition

The Hobbs algorithm is a historical, syntax-based method for finding candidate antecedents of pronouns in text. It searches constituency parse trees in a specified order and checks whether candidate noun phrases are compatible with the pronoun. The name usually refers to the simpler tree-search approach in Jerry Hobbs's 1978 paper, rather than the paper's broader semantic interpretation approach.

What it means

An antecedent is an expression that a pronoun refers back to, such as “the robot” for “it” in a short passage. Finding that link helps a system connect information across sentences. Hobbs's method uses grammatical structure to decide which candidates to consider first.

This makes it different from simply choosing the nearest noun. A candidate's position in a parse tree affects when it is examined, and an incompatible candidate can be rejected. Nevertheless, a syntactically preferred candidate is not necessarily the speaker's intended referent. Meaning and discourse can overturn a plausible structural guess.

How it works

The method begins at the noun-phrase node containing the pronoun and moves upward through relevant phrase or sentence nodes. It explores prescribed portions of the tree, proposing noun phrases in an ordered search. If appropriate, it searches earlier sentence trees, beginning with the most recent, using left-to-right breadth-first traversal.

Breadth-first means examining shallower nodes before deeper ones. Compatibility checks, including number and gender in the historical account, filter proposals. Detailed behavior depends on the assumed grammatical representation, so an implementation must specify its trees and constraints.

For practical inspection, preserve the parse, candidate order, rejection reasons, and selected link. This trace lets a reviewer distinguish a search-order decision from a parsing error or a compatibility error. A modern adaptation must also handle contemporary pronoun use and avoid treating guessed personal attributes as verified facts. A short glossary explanation is not a complete implementation specification.

Three practical examples

1. NLP education: Students trace a pronoun through a small constituency tree to see how syntactic structure can guide reference resolution without a trained neural ranking model.

2. Baseline comparison: A research team uses a carefully specified implementation as a historical baseline, comparing errors with another resolver on the same annotated passages.

3. Error analysis: An engineer investigates why a document-processing prototype linked a pronoun to the wrong entity. The search trace reveals whether the intended candidate was absent, rejected, or encountered too late.

Think of it this way

Think of searching an organizational chart according to a fixed route rather than scanning a list of names by distance on a page. The route determines which candidates appear first. The analogy captures ordered structural search, but a sentence tree represents grammar rather than authority, and finding the first eligible entry does not prove that it is the intended referent.

Worked example

Use the invented passage “The robot moved the crates. It stopped.” Assume a simple constituency analysis: the first sentence has subject NP “the robot” and a verb phrase containing object NP “the crates”; the second has subject NP “it” and a verb phrase “stopped.” Treat robot as compatible with singular it and crates as plural.

Starting at the pronoun's NP, the search reaches the second sentence's top sentence node. There is no earlier candidate on the required leftward branch in that sentence. Searching the preceding sentence breadth-first encounters its subject NP before the deeper object NP. “The robot” is compatible and is selected.

This outcome differs from a nearest-noun shortcut, which would first encounter crates in the preceding text. If the second sentence instead begins “They,” and the exercise explicitly uses plural they, robot fails the number check and crates can be proposed later. This restricted exercise does not deny singular they or specify every case of the original algorithm. It demonstrates how tree order and declared compatibility assumptions interact in a simple intersentence example.

Illustrative case study

Fictional case: A technical publisher tests a reference-resolution prototype on maintenance instructions. Some passages describe a machine and several components, then use “it” in the next sentence.

The team compares proposed links against editor annotations and reviews disagreements. In one passage, the tree-search preference selects the machine, while the surrounding instructions make a component the intended referent. The team records this as a semantic-context limitation rather than changing the reference answer to agree with the algorithm. For publication, ambiguous source instructions are rewritten where appropriate. This case illustrates review practice, not a measured benchmark result.

Common misconceptions and mistakes

Do not reduce the Hobbs algorithm to “pick the closest noun” or “always pick the subject.” Its proposal order follows a structured procedure, and compatibility constraints matter. Conversely, do not assume a detailed parse is sufficient for every reference problem.

Another mistake is presenting a simplified textbook variant as the exact original algorithm. Variants can be useful for teaching, but their search order and coverage must be labeled. Also distinguish pronoun resolution from every form of coreference: matching two names, resolving an omitted argument, and interpreting a pronoun that refers to an entire event can require different treatment.

Business use

For business systems, the main value is an understandable baseline and a clear illustration of reference errors. A wrong link can attach an action or attribute to the wrong item in an extracted record. Evaluate the resulting entity links and downstream records, not just whether the algorithm runs. Preserve uncertain cases for review when the source wording itself supports several interpretations.

Check your understanding

Why is robot considered before crates in the worked example?

Answer: In the assumed tree, the subject NP is shallower than the object NP, so the preceding-sentence breadth-first search reaches it first.

Related terms

Antecedent
Anaphora
Breadth First Search
Pronoun resolution
coreference
constituency parse
agreement
discourse context

Learn next

Draw the two sentence trees in the example and mark the order of candidate visits. Then compare a syntactic preference with a passage whose intended reference requires event knowledge. Study modern mention-ranking systems while retaining explicit tests for ambiguous references.

Sources

Primary source: Jerry R. Hobbs, Resolving Pronoun References (1978), university-hosted reproduction: https://personal.utdallas.edu/~sanda/courses/Discourse/Papers/Hobbs_resolving_pronouns.pdf . The tree-search procedure and surrounding assumptions were read. The robot passage is an original restricted example, not a complete algorithm trace for all constructions. Historical grammatical assumptions require explicit review in modern implementations.

Previous: Ho - Kashyap algorithm Next: Hodrick - Prescott filter