Language models, like Large Language Models (LLMs), have significantly advanced text generation for different domains, including error correction, text simplification, paraphrasing, and style transfer. These models can generalize tasks, reducing the need for many exemplars through fine-tuning with instructions. Text editing benchmarks can overcome the challenge in fine-tuning text editing models due to factors including lack of task-specific datasets.
A recent development in this field is CoEdIT, an AI-based text editing system introduced by researchers from Grammarly and the University of Minnesota. The system can add, delete, or change words, phrases, and sentences, delivering top-notch performance on various text editing benchmarks. The research team demonstrated that CoEdIT could perform modifications even for unseen, adjacent, and composite instructions, concluding that it can meaningfully assist authors in text rewriting.
The research made three notable contributions. Firstly, the team achieved state-of-the-art results on three stylistic editing tasks and others, significantly exceeding other models in quality and performance. Secondly, their smallest model outperformed other text editing and instruction-tuned models. Lastly, the data and models are available for public use, and CoEdIT performed well on novel, neighboring tasks not explored during the fine-tuning stage.
The researchers sought to determine if CoEdIT could adhere to editing guidelines, perform edits for novel instructions effectively, and aid authors to write more efficiently. They evaluated other text editing LLMs adapted using instruction-specific data, comparing them to their main alternatives including FLAN-T5 models and others that showed superior performance on text editing tasks.
Their comparisons encompassed several subsets, including PEER, GPT3, Alpaca, and others. In their several assessments, CoEdIT performed better than larger models like ChatGPT and InstructGPT and others of comparable size. This indicates that densifying task/instruction space may be more beneficial than scaling model size as the former models are underfitted.
Though CoEdIT scored impressively across several text editing benchmarks, limitations such as being primarily sentence-level oriented restricts its efficacy on longer real-world text sequences. Moreover, its primary focus is on non-meaning-altering text alterations. Future work may focus on addressing these limitations.
In conclusion, Language models like LLMs have significantly advanced the generation of coherent text with different modifications like error correction, simplification, and style transfer. The introduction of AI-based systems like CoEdIT promises advancements in text editing, significantly aiding authors in text rewriting amid limitations like predominantly sentence-level orientation and an exclusive focus on non-meaning-altering text changes. Future work may focus on overcoming these limitations for even better performance.