CompactIE: Compact Facts in Open Information Extraction

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 900 - 910 July 10-15, 2022 ©2022 Association for Computational Linguistics CompactIE: Compact Facts in Open Information Extraction Farima Fatahi Bayat University of Michigan farimaf@umich.edu Nikita Bhutani Megagon Labs nikita@megagon.ai H. V. Jagadish University of Michigan jag@umich.edu Abstract A major drawback of modern neural OpenIE systems and benchmarks is that they prioritize high coverage of information in extractions over compactness of their constituents. This severely limits the usefulness of OpenIE ex- tractions in many downstream tasks. The util- ity of extractions can be improved if extrac- tions are compact and share constituents. To this end, we study the problem of identifying compact extractions with neural-based meth- ods. We propose COMPACTIE, an OpenIE sys- tem that uses a novel pipelined approach to produce compact extractions with overlapping constituents. It first detects constituents of the extractions and then links them to build extrac- tions. We train our system on compact extrac- tions obtained by processing existing bench- marks. Our experiments on CaRB and Wire57 datasets indicate that COMPACTIE finds 1.5x- 2x more compact extractions than previous systems, with high precision, establishing a new state-of-the-art performance in OpenIE. 1 Introduction A popular domain-agnostic paradigm to structure the raw text is open information extraction (Ope- nIE) (Banko et al., 2007). Not relying on any pre- defined schema, OpenIE systems typically extract information as (subject; relation; object) triples. The extracted information is then used in sev- eral downstream applications, including answering questions (Khot et al., 2017), summarizing docu- ments (Hao et al., 2018; Ji et al., 2013), and popu- lating knowledge bases (Fan et al., 2019). Despite much progress, state-of-the-art neural OpenIE systems focus on covering more informa- tion from the input sentence often at the cost of util- ity and compactness of the extracted triples. The ex- tracted triples have long, over-specific constituents (i.e. the relation and its arguments). Figure 1 illus- trates such example triples produced by a popular Beth is the second child of Henry, born in wedlock. IMoJIE Extractions E1: ( Beth ; is ; the second child of Henry born in wedlock ) Compact Extractions E1: ( Beth ; is ; the second child of Henry ) E2: ( the second child of Henry ; born ; in wedlock ) The rest of the group reach a small shop , where the crocodile breaks through a wall and devours Annabelle. IMoJIE Extractions E1: ( The rest of the group ; reach ; a small shop , where the crocodile breaks through a wall and devours Annabelle ) E2: ( the crocodile ; devours ; Annabelle a small shop) Compact Extractions E1: ( The rest of the group ; reach ; a small shop ) E2: ( crocodile ; breaks ; through a wall ) E3: ( crocodile ; devours ; Annabelle ) Figure 1: Example sentences with non-compact triples from IMoJIE vs. compact triples from our benchmark. Compact triples can …

Content truncated to 500 words