# An Effective System for Multi-format Information Extraction

**Authors:** Yaduo Liu, Longhui Zhang, Shujuan Yin, Xiaofeng Zhao, Feiliang Ren  
**arXiv:** [2108.06957](https://arxiv.org/abs/2108.06957)  
**Date:** 2021-08-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.06957v1/

## Key points

- System for LIC-2021 multi-format Information Extraction (IE) task
- Evaluation of information extraction from multiple dimensions
- Multiple slots relation extraction
- Event extraction at sentence-level and document-level
- Methods employed to address challenges in the competition
- Schema disintegration method for relation extraction subtask
- Voting-based method for maximizing model utilization
- Conversion of sentence-level event extraction into Named Entity Recognition (NER) task
- Pointer labeling based approach for efficient event extraction
- Auxiliary trigger recognition model for aiding event extraction
- Integration of trigger features using multi-task learning mechanism
- Encoder-Decoder based method with Transformer-alike decoder architecture for document-level event extraction subtask
- Relation extraction: F1 score of 79.887%
- Sentence-level event extractions: F1 score of 85.179%
- Document level event extractions: F1 score of 70.828%
- Unannotated triples negatively impacting performance in relation extraction
- Challenges in processing long text in document-level event extraction subtask
- Correctly extracting two arguments of one event when they are far apart in a sentence or document requires further study.
- Funding support from various sources.

## Summary

This paper presents our system for the LIC-2021 multi-format Information Extraction (IE) task. The task aims to evaluate information extraction from various dimensions, including multiple slots relation extraction and event extraction at both sentence-level and document-level. To address the challenges in this competition, we employ different methods. For the relation extraction subtask, we tackle the issue of multiple-O-values schema by using a schema disintegration method. This helps in converting the subtask into a traditional triple extraction task. Additionally, we design a voting-based method that maximizes the utilization of existing models. For the sentence-level event extraction subtask, we convert it into a Named Entity Recognition (NER) task. We utilize a pointer labeling based approach for efficient event extraction. Furthermore, recognizing that annotated trigger information can aid in event extraction, we develop an auxiliary trigger recognition model. We integrate trigger features into the event extraction model using multi-task learning mechanism. In order to handle document-level event extraction subtask, we propose an Encoder-Decoder based method with a Transformer-alike decoder architecture. Our system achieves promising results and ranks No.4 on the test set leaderboard of this multi-format IE task with F1 scores obtained for relation extraction, sentence-level event extractions and document level event extractions being 79.887%, 85.179% and 70.828% respectively. However there is still room for improvement in our system as many triples are not annotated which negatively impacts performance and processing long text remains challenging in document level event extraction subtask along with extracting two arguments of one event correctly when they are far apart in either a sentence or a document being an area that requires further study .In conclusion , our system demonstrates effectiveness in addressing various challenges posed by the LIC 2021 multi format IE task and achieves competitive performance while there are opportunities for further exploration and improvement in future research .This work is supported by National Key R&D Program of China (No .2018YFC0830701), National Natural Science Foundation of China (No .61572120), Fundamental Research Funds for Central Universities (No .N181602013 & N171602003), Ten Thousand Talent Program (No .ZX20200035) & Liaoning Distinguished Professor (No .XLYC1902057).