OpenIE Most OpenIE systems produce predicate-argument structures that resemble PredPatt output at first glance.
Researchers at the University of Washington, home of related systems such as TextRunner (Yates et al., 2007) write:
"SRL [semantic role labeling] and Open IE are quite related [...] semantically labeled arguments correspond to the arguments in Open IE extractions, and verbs often match up with Open IE relations" (Christensen et al., 2011).
The most closely related are those relying on dependency trees, including ClausIE (Del Corro et al., 2013) and ArgOE (Gamallo & Garcia, 2015), which support multiple languages (Spanish, Portuguese, Galician and English), but are not based on UD. A key distinction is the flat representation of such OpenIE statements: relational triples, with predicates and arguments all considered strings with no internal (compositional) structure. Further, the sets of rules employed by those systems are less complex, can be dependent on word order, and rely on lexicalization (language-specific). For example, consider the forced binary outputs:
John gave the book to Mary yesterday.
ClausIE:
⇒ [John] gave [the book to Mary]
⇒ [John] gave [the book yesterday]
⇒ [John] gave [the book]
PredPatt:
⇒ [John] gave [the book] to [Mary] [yesterday]
Angeli et al. (2015) developed an OpenIE system atop the Stanford Depedencies, using knowledge base completion as an end-task to derive distant supervision for a machine learning approach. They map open-domain relations extracted to closed-domain relations using association mining techniques to improve the coverage of a knowledge base population system. The bias towards the knowledge base completion end-task leads to spurious extractions and low recall, and is not geared toward succinct, human interpretable, language-neutral rules.
PropBank PredPatt extractions are similar to those found in the manually constructed PropBank (Palmer et al., 2005), without the attendant semantic roles. PredPatt extracts more types of structures, e.g., copula, appositives, possessives. There are also a number of differences in argument boundaries.
Textual Inference observe that many shallow approaches to recognizing textual inference break down as the text and hypothesis become lengthier and more complex. Their solution is to extract a set of discourse commitments from the text and hypothesis, and determine entailment by searching for matches over these resulting sets of simpler propositions. In this case, the process of extracting discourse commitments from text could be aided by a tool like PredPatt.
Meaning representations PredPatt is not a semantic representation, but might be a useful preprocessing step towards building one. Richer semantic labels could be assigned in later phases, e.g., PropBank roles (Palmer et al., 2005), or proto-roles (Reisinger et al., 2015). Unlike the recently popular Abstract Meaning Representation (Banarescu et al., 2013), PredPatt retains a strict commitment to the syntax/semantics interface. This is both to directly benefit from ongoing improvements by the community in syntactic parsing, as well as to enable linguistically well-motivated compositional analysis. These goals likewise motivated the KNEXT effort (Schubert, 2002; Van Durme & Schubert, 2008), a transformation process that was English-specific, applied to constituency parses to derive shallow logical forms. A similar approach was explored in Rudinger & Van Durme (2014) with the goal of understanding the extent to which the Stanford dependency syntax representation is a semantic representation.