Source-linked AI summary
Graphene: Semantically-Linked Propositions in Open Information Extraction
Matthias Cetto, Christina Niklaus, André Freitas, Siegfried Handschuh
TL;DR
Open IE systems struggle with complex sentences that hide relations or combine multiple facts in oversized arguments. Graphene applies clausal and phrasal disembedding with rhetorical relation identification to produce hierarchical core and contextual propositions. Its reference implementation achieves the highest average precision in the benchmark, while the transformation process improves the AUC of existing systems by up to 63%.
Problem
Existing Open IE systems can miss relations in complex constructions or produce over-specified arguments that combine semantically independent facts.
Method
Graphene recursively simplifies complex sentences through clausal and phrasal disembedding, then represents core facts and contextual information linked by rhetorical relations.
Results
50.1% average precision makes Graphene the best-performing system, while preprocessing improves existing systems' AUC by up to 63%.
Takeaways & Limitations
The two-layered representation preserves semantic connections while enriching Open IE relations without losing precision or recall.
Abstract
from arXiv · showhide
We present an Open Information Extraction (IE) approach that uses a two-layered transformation stage consisting of a clausal disembedding layer and a phrasal disembedding layer, together with rhetorical relation identification. In that way, we convert sentences that present a complex linguistic structure into simplified, syntactically sound sentences, from which we can extract propositions that are represented in a two-layered hierarchy in the form of core relational tuples and accompanying contextual information which are semantically linked via rhetorical relations. In a comparative evaluation, we demonstrate that our reference implementation Graphene outperforms state-of-the-art Open IE systems in the construction of correct n-ary predicate-argument structures. Moreover, we show that existing Open IE approaches can benefit from the transformation process of our framework.
1 Introduction
Existing Open IE systems either miss relations in complex constructions or produce over-specified propositions that mix independent facts. Graphene addresses both problems by simplifying sentences into core facts with semantically linked contextual information.
- 1 Introduction: Complex constructions can cause OLLIE and ClausIE to miss asserted relations, such as Sonia Sotomayor becoming the first Supreme Court Justice of Hispanic descent.
- 1 Introduction: Long argument phrases can combine multiple semantically independent facts, making extractions difficult for downstream applications such as question answering and textual entailment.ClausIE incorporates such information into over-specified argument components.
- 1 Introduction: Graphene uses clausal and phrasal disembedding to transform complex sentences into clean, compact canonical structures that retain the main information.The transformation removes non-central clauses and phrases and converts them into independent sentences.
- 1 Introduction: Its two-layered representation separates core relational tuples from contextual arguments and links them through rhetorical relations, preserving connections between propositions.Graphene distinguishes simple textual contextual arguments from arguments linking to other propositions.
2 Related Work
Open IE research includes learning-based, rule-based, and clause-based approaches, with clause-based systems restructuring complex sentences before extraction. Graphene extends this restructuring direction through its own transformation pipeline.
- 2 Related Work: Learning-based systems acquire extraction patterns from heuristic or external sources, using features such as part-of-speech tags, noun-phrase chunks, or dependency parses.TEXTRUNNER, WOE, and OLLIE exemplify this line of work.
- 2 Related Work: Rule-based systems use hand-crafted lexical, syntactic, or dependency rules, including approaches designed to extract complete n-ary facts.REVERB reduces incoherent or overspecified relation phrases, while KRAKEN and EXEMPLAR target arbitrary-arity relations.
- 2 Related Work: Clause-based approaches restructure complex sentences into independent clauses that are easier to segment into Open IE tuples.ClausIE derives coherent clauses from dependency relations before generating predicate-argument extractions.
- 2 Related Work: Graphene builds on sentence-rearrangement strategies by recursively transforming complex inputs into simpler sentences with accompanying contextual information.
3 Proposed Open IE Approach
Graphene transforms complex sentences into compact, syntactically sound sentences organized in a discourse tree, then extracts propositions linked by rhetorical and contextual relations.
- 3.1 Transformation Stage: Graphene recursively disembeds clausal and phrasal constituents to transform complex sentences into simpler, compact, syntactically sound sentences with contextual information.The transformation stage targets both complex nested structures and supplementary phrases, producing canonical sentence forms where needed.
- 3.1.3 Transformation Rules: Constituency labels and cue phrases help identify rhetorical relations such as Condition, Enablement, or Background between the generated sentences.Syntactic features come from phrasal composition, while lexical features include discourse cue phrases such as “because,” “after that,” and “in order to.”
- 3.1.1 Concept of the Discourse Tree Creation: The discourse tree represents simplified sentences hierarchically, distinguishing core coordinate sentences from subordinate context sentences and connecting them with rhetorical relations.The framework adapts RST concepts while generating the final simplified sentences during top-down transformation rather than identifying fixed EDUs beforehand.
- 3.1.2 Algorithm: A top-down traversal applies the first matching rule from 16 ordered hand-crafted syntactic patterns, recursively replacing matched leaves with two or more simplified sentence leaves.Rules use syntactic and lexical features from phrase-structure parses, including Tregex patterns, and may paraphrase extracted material into grammatical sentences.
- 3.2 Relation Extraction: Graphene feeds each generated leaf sentence to an Open IE extractor, identifies its representative proposition by head-verb heuristics, and transfers discourse relations as semantically typed arguments to form n-ary relations.A representative proposition contains the input sentence’s head verb in its relation phrase or as its object argument; Graphene’s baseline uses a subject-predicate-object extraction pattern.
4 Evaluation
Graphene is evaluated against Open IE systems on a QA-SRL benchmark, both as a reference implementation and as a transformation framework for other extractors. It achieves the highest average precision, while disembedding improves the AUC of most tested systems.
- Evaluation Setup: The benchmark contains more than 10,000 extractions from 3,200 Wikipedia and Wall Street Journal sentences and supports precision-recall comparison.Graphene is evaluated as a reference implementation, while other systems are tested on raw and simplified inputs.
- Results and Discussion: 50.1% average precision makes Graphene the best-performing system, ahead of OpenIE-4 at 44.6% and PropS at 42.4%.Graphene’s recall is 27.2%, comparable to PropS at 26.7%, but below ClausIE at 33.0% and OpenIE-4 at 32.5%.
- Results and Discussion: Graphene’s 13.6% AUC trails OpenIE-4 at 14.5% but exceeds PropS at 11.3% and ClausIE at 9.3%.These results are reported as overall Area Under the Curve scores within the benchmark framework.
- Results and Discussion: Disembedding increases AUC for every tested system except OpenIE-4, whose AUC decreases by 18%.Stanford Open IE gains 63%, while OLLIE and REVERB gain 34% and 22%, respectively.
- Results and Discussion: The framework improves OLLIE and REVERB mainly through recall gains, whereas ClausIE mainly benefits through a precision increase.OLLIE and REVERB gain recall by 4% and 26%; ClausIE gains precision by 16%.
5 Conclusion
The paper presents Graphene as a two-layered Open IE approach that separates core facts from contextual information while preserving their rhetorical links. Graphene achieves the best average precision in the benchmark, and the transformation process improves existing systems’ AUC by up to 63%.
- Conclusion: Graphene represents simplified sentences as core facts with accompanying contexts linked by rhetorical relations, preserving connections needed to reconstruct input information.Its transformation uses clausal and phrasal disembedding to detach contextual information from core propositions and create more compact structures.
- Conclusion: 50.1% average precision makes Graphene the benchmark’s best system, while its 27.2% recall is comparable to other high-precision systems.The results support semantically enriched relational tuples without losing precision or recall.
- Conclusion: Clausal and phrasal disembedding improves the AUC of state-of-the-art Open IE systems by up to 63%.The process converts contextual information into additional arguments and avoids overspecified argument phrases.