In the quest to turn unstructured text into structured knowledge, Natural Language Processing (NLP) has long suffered from a "verb-centric" bias. Traditional Information Extraction (IE) systems are designed to hunt for actions: they look for a subject, a verb, and an object (S-V-O). If a sentence says "Apple acquired Beats," the system easily identifies the relationship.
However, a vast amount of human knowledge is never expressed through verbs. Facts are frequently buried in noun phrases, appositives, and compound descriptions. When we read "Apple CEO Tim Cook," we instantly know the relationship between the person and the company, even though no verb links them. Researchers are now shifting focus toward nominal relation extraction—a direction that treats nouns, not just verbs, as the primary anchors of semantic information.
The Verb Centric Bias
For decades, the standard pipeline for Open Information Extraction (Open IE) assumed that a "relation" must be mediated by a predicate. As Mausam notes in Open Information Extraction Systems and Downstream Applications, early systems focused almost exclusively on verb-based patterns. This approach works for simple sentences but fails on the "long tail" of linguistic complexity.
This bias creates a significant gap in data coverage. Gabriel Stanovsky and Ido Dagan argue in Creating a Large Benchmark for Open Information Extraction that even our evaluation frameworks are flawed; by focusing on verbal predicates, we fail to measure how well systems capture relations expressed through nominalizations (e.g., "the acquisition of Beats by Apple"). If the extraction logic is biased toward specific syntactic heads, it misses patterns centered on noun phrases or prepositional constructs De Saeger et al..
Why Nouns Matter for Knowledge Bases
The push toward nominal extraction isn't just a linguistic curiosity—it is a practical necessity for building Knowledge Bases (KBs). In ReNoun: Fact Extraction for Nominal Attributes, researchers point out a fundamental mismatch: the vast majority of attributes in structured databases (like "job title" or "nationality") are represented as nouns.
Relying on verbs to find these facts is an indirect and error-prone detour. For example, a system might wait to find the sentence "Tim Cook is the CEO of Apple" to extract a fact. But Stephen Soderland and Oren Etzioni demonstrate in Open Information Extraction to KBP Relations in 3 Hours that facts are more often found in "static" descriptions or parenthetical phrases. By ignoring appositives and noun modifiers, systems overlook the most efficient path to populating a KB Gilmer and Soderland.
Mechanisms of Nominal Extraction
To capture these elusive relations, researchers have developed several distinct strategies:
- Noun-Mediated Triggers: Instead of starting with a verb list, systems like EXEMPLAR use dependency parsing to identify nouns that act as "triggers" for relations de Sá Mesquita et al.. This allows the system to treat "CEO" or "founder" as the functional equivalent of a verb.
- N-ary Relation Mapping: Real-world events are often complex, involving multiple participants. Feiyu Xu and Hans Uszkoreit show that expanding patterns to include noun phrases significantly improves the recall of these n-ary relations, capturing roles that a simple S-V-O triple would truncate A Seed-driven Bottom-up Machine Learning Framework for Extracting Relations of Various Complexity.
- Noun Phrase (NP) Enrichment: Understanding a text requires recognizing the "dense network" of relations between entities that aren't explicitly linked. Yanai Elazar and Yoav Goldberg propose "NP Enrichment" to capture these implicit connections, such as the part-whole relationship between a "room" and its "window" Text-based NP Enrichment.
Beyond English: The Structural Skeleton
The move away from verb-centrism also makes IE systems more portable across languages. In Porting an Open Information Extraction System from English to German, Stanovsky et al. found that while verb conjugations and placements vary wildly between languages, the semantic relationships within noun phrases are often more stable.
By focusing on the "structural skeleton" of language—using tools like SMiLe-OIE to incorporate syntactic multi-view learning Dong et al.—modern models can identify relations even when a part-of-speech tagger fails to find a clear predicate. As Christina Niklaus concludes in her survey, incorporating patterns for nouns and adjectives doesn't just add a few extra facts; it fundamentally increases the "total yield" of structured data from diverse text sources.