HermEs: Interactive Spreadsheet Formula Prediction via Hierarchical Formulet Expansion

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8356–8372 July 9-14, 2023 ©2023 Association for Computational Linguistics HERMES: Interactive Spreadsheet Formula Prediction via Hierarchical Formulet Expansion Wanrong He1∗, Haoyu Dong2†, Yihuai Gao1, Zhichao Fan2, Xingzhuo Guo1, Zhitao Hou2, Xiao Lv2, Ran Jia2, Shi Han2, Dongmei Zhang2 1Tsinghua University, 2Microsoft Research Asia {hwr19,gao-yh18,gxz19}@mails.tsinghua.edu.cnn, {hadong,v-zhichaofan,zhith,xilv,raji,shihan,dongmeiz}@microsoft.com Abstract We propose HERMES, the first approach for spreadsheet formula prediction via HiEraRchi- cal forMulet ExpanSion, where hierarchical ex- pansion means generating formulas following the underlying parse tree structure, and For- mulet refers to commonly-used multi-level pat- terns mined from real formula parse trees. HER- MES improves the formula prediction accuracy by (1) guaranteeing correct grammar by hierar- chical generation rather than left-to-right gen- eration and (2) significantly streamlining the token-level decoding with high-level Formulet. Notably, instead of generating formulas in a pre-defined fixed order, we propose a novel sampling strategy to systematically exploit a variety of hierarchical and multi-level expan- sion orders and provided solid mathematical proof, with the aim of meeting diverse human needs of the formula writing order in real ap- plications. We further develop an interactive formula completion interface based on HER- MES, which shows a new user experience in https://github.com/formulet/HERMES. 1 Introduction Hundreds of millions of people use spreadsheets, such as Excel and Google Sheets, for data storage and management. The semi-structured context in- fers potential relations among data in cells. Users can add formulas in their spreadsheets to process and analyze data based on such relations. Although spreadsheet formula language targets general users and is much simpler than programming languages like C++, it is still difficult for most Excel users without any programming experience to master. To write a formula for a cell, users need to find appro- priate formula functions, such as IF and SUBTOTAL, compose them with correct grammar, and fill in proper values and cell references. This process could be time-consuming and error-prone. ∗Work done during Wanrong’s internship at Microsoft Research Asia. †Corresponding author. For computers, spreadsheet formula prediction is also a challenging task, which requires under- standing both textual and numerical data, diverse table structures, and relationships between cells. Thanks to the advances in language model (LM) pretraining, natural language (NL) understanding, and spreadsheet table understanding and reasoning, researchers have developed systems that could pre- dict the formula of a selected cell in table (Chen et al., 2021; Cheng et al., 2022). Those systems transform users from “writers” into “decision mak- ers”, greatly reducing the required efforts for writ- ing formulas. We have to argue, though, directly applying autoregressive sequence generation to for- mula generation has fundamental weaknesses. First, they fail to utilize the inner structures in spreadsheet formulas. Spreadsheet formulas have well-defined grammar and structure, including the usage of operators, parenthesis, constant values, cell references, and their combinations. Without taking the structure into consideration, a minor er- ror made where the model is uncertain during the left-to-right decoding process might cause subse- quent grammar mistakes (Guo et …

Content truncated to 500 words