Overcoming the Verb-Centric Bias in Information Extraction

64 concepts48 papers1 proposal

If you ask a traditional Natural Language Processing (NLP) system to extract facts from the sentence "Google acquired YouTube," it succeeds effortlessly. The verb "acquired" serves as an explicit, structural anchor connecting the subject to the object, yielding a clean relational triple: (Google, acquired, YouTube).

But human language rarely restricts itself to such neat, action-oriented structures. Consider alternative phrasings: "Google's acquisition of YouTube..." or "Google CEO Sundar Pichai announced..." In these cases, the vital relationships—an acquisition and a job title—are embedded entirely within noun phrases, modifiers, and appositives. No central verb links them.

Historically, Open Information Extraction (Open IE) and relation extraction systems suffered from a severe verb-centric extraction bias. By assuming that relationships are always anchored by verbs, early frameworks systematically ignored the vast wealth of information conveyed through non-verbal structures. Overcoming this bias by designing systems capable of extracting nominal and non-verbal relations has become a crucial frontier in building comprehensive knowledge bases.

The Blind Spot of Verb-Centric Systems

The historical reliance on verbs was a matter of convenience. Early Open IE systems relied heavily on identifying explicit predicate-argument structures where the predicate was almost always a verb. As Mausam notes in a ten-year retrospective on Open IE, this narrow focus meant that early approaches completely missed information conveyed through noun phrases and appositives, resulting in highly incomplete knowledge acquisition.

This limitation is more than a minor technical omission; it represents a fundamental bottleneck in data yield. In a comprehensive survey of the field, Christina Niklaus et al. point out that restricting extraction to verb-mediated relations severely caps the total volume of structured data an algorithm can harvest. Language is dense with static descriptions, parentheticals, and compound nouns that contain facts completely invisible to a verb-hunting parser.

The Knowledge Base Mismatch

The consequences of this bias become starkly apparent when looking at the ultimate destination for extracted data: structured knowledge bases (KBs).

Mohamed Yahya et al. identify a fundamental irony in ReNoun: Fact Extraction for Nominal Attributes: the vast majority of attributes stored in real-world knowledge bases are inherently nominal. Properties like CEO, spouse, profession, or birthplace are nouns, not actions.

Verb-Centric Approach (Indirect & Error-Prone):
"Sundar Pichai serves as the CEO of Google" ---> [Verb: serves as] ---> (Sundar Pichai, is CEO of, Google)

Nominal-Centric Approach (Direct & Efficient):
"Google CEO Sundar Pichai" -------------------> [Noun Trigger: CEO] -> (Sundar Pichai, CEO, Google)

When an extraction system is biased toward verbs, it must rely on finding an indirect verbal intermediary (such as "X serves as the CEO of Y") to extract a noun-based attribute. This indirect step introduces errors and completely misses niche attributes that are naturally expressed through compact noun phrases. Filipe de Sá Mesquita et al. argue that because a massive portion of relational data is conveyed through nouns, systems that fail to recognize nouns as valid relation indicators miss substantial amounts of data, especially in concise or diverse text types.

Anatomy of Non-Verbal Relations

To capture the information missed by verb-centric models, researchers have isolated several distinct linguistic mechanisms that express facts without active predicates:

  • Nominalizations: These are nouns derived from verbs (e.g., "destruction", "acquisition", "marriage"). Gabriel Stanovsky et al. observe in Supervised Open Information Extraction that extracting structured propositions from nominalizations is significantly more challenging than standard verbal extraction. Nominalizations lack the rigid, predictable argument structures that verbs provide, making them difficult to parse even for supervised deep learning models.
  • Appositions and Titles: Information is frequently packed into parentheticals or titles preceding a name (e.g., "biologist Jane Cooke"). Kiril Gashteovski et al. demonstrate with their system, MinIE, that these implicit syntactic structures can be systematically unpacked to generate "minimal" facts. For instance, a title preceding a name can be automatically converted into a structured triple representing a role or identity relationship, capturing data embedded entirely within a noun phrase.
  • Modifiers and Adjectives: In concise text, a majority of assertions are made through non-verbal predicates like adjectives and compound nouns. Rachel Wities et al. show that capturing these implicit relationships is essential for building a comprehensive open knowledge representation. Without this capability, as Feiyu Xu 0001 et al. argue, complex n-ary relations—which describe real-world events involving multiple participants and attributes but lack a central predicate—are entirely lost.

Shifting the Extraction Paradigm

How do modern systems overcome the verb-centric bias? The transition requires moving away from rigid verb-template matching toward flexible, noun-aware architectures.

One foundational approach is to explicitly expand rule-based systems to include noun-mediated triggers. For example, the EXEMPLAR framework, introduced by Filipe de Sá Mesquita et al., incorporates noun-based triggers alongside traditional verb-based triggers. By recognizing nouns as valid relation indicators, the system captures a broader range of semantic information from both well-formed and ungrammatical text without incurring the heavy computational cost of deep semantic parsing.

However, scaling these rules across languages introduces significant engineering hurdles. Gabriel Stanovsky et al. note that when porting extraction systems from English to languages like German, capturing semantic relations where a noun serves as the primary link requires specialized logic that differs fundamentally from standard verb-centric extraction rules.

To bypass the limitations of hand-crafted rules, modern architectures leverage deep learning to learn non-verbal patterns implicitly. Systems like Multi$^2$OIE, developed by Youngbin Ro et al., utilize multi-head attention mechanisms and contextual language models like BERT. Because these models look at the entire context of a sentence simultaneously rather than anchoring on specific part-of-speech tags, they can successfully extract nominal relations and complex predicates where traditional sequence-labeling models fail.

By recognizing that verbs are merely one of many vehicles for meaning, the shift toward non-verbal relation extraction allows information extraction systems to align more closely with the true structure of human language and the practical needs of structured knowledge bases.

Key Foundational Work

Book Reviews (2002) — contributed 1 concept elements to this direction.
An Improved Extraction Pattern Representation Model for Automatic IE Pattern Acquisition (2003) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_verb_centric_extraction_bias_ababde59,
  title = {verb-centric extraction bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/verb-centric-extraction-bias/rdr_ababde5912eb47b88310a87f4cf78f3d},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_verb_centric_extraction_bias_ababde59_v1,
  title = {verb-centric extraction bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/verb-centric-extraction-bias/rdr_ababde5912eb47b88310a87f4cf78f3d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake