# Multitasking Framework for Unsupervised Simple Definition Generation

**Authors:** Cunliang Kong, Yun Chen, Hengyuan Zhang, Liner Yang, Erhong Yang  
**arXiv:** [2203.12926](https://arxiv.org/abs/2203.12926)  
**Date:** 2022-03-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.12926v1/

## Key points

- Authors propose Simple Definition Generation (SDG) task to assist language learners and individuals with low literacy levels
- Lack of learner's dictionaries in many languages poses a challenge for SDG
- Introduce multitasking framework called SimpDefiner using standard dictionary and corpus of simple texts
- Parameter sharing scheme between decoders enables generation of complex and simple definitions simultaneously
- Incorporate text reconstruction task to control complexity and language modeling task to enhance performance
- Evaluation on novel test set in English aligning Oxford Dictionary (OD) and Oxford Advanced Learner's Dictionary (OALD)
- SimpDefiner outperforms other models in generating accurate and straightforward definitions
- Future aim to combine method with prompt learning techniques for further complexity conditioning
- Research accepted by ACL 2022 for presentation at main conference

## Summary

The authors propose a novel task of Simple Definition Generation (SDG) to assist language learners and individuals with low literacy levels. The lack of learner's dictionaries in many languages poses a significant challenge for this task. To address this issue, the authors introduce a multitasking framework called SimpDefiner that utilizes a standard dictionary with complex definitions and a corpus of simple texts. By carefully designing a parameter sharing scheme between two decoders, the framework can generate both complex and simple definitions simultaneously. Additionally, the authors incorporate a text reconstruction task to control text complexity and a language modeling task to enhance the decoder's performance. To evaluate the effectiveness of their proposed framework, the authors construct a novel test set in English by aligning two dictionaries - Oxford Dictionary (OD) and Oxford Advanced Learner's Dictionary (OALD). Automatic and manual evaluations demonstrate that the SimpDefiner framework outperforms other models and generation-simplification pipelines in terms of generating accurate and straightforward definitions. Moving forward, the authors aim to explore combining their current method with prompt learning techniques to allow users to condition the complexity of generated definitions further. This research was accepted by ACL 2022 for presentation at its main conference.