Students Who Study Together Learn Better: On the Importance of Collective Knowledge Distillation for Domain Transfer in Fact Verification

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 719–733 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 719 Truth-Conditional Captioning of Time Series Data Harsh Jhamtani School of Computer Science Carnegie Mellon University jharsh@cs.cmu.edu Taylor Berg-Kirkpatrick Computer Science and Engineering University of California San Diego tberg@ucsd.eng.edu Abstract In this paper, we explore the task of automati- cally generating natural language descriptions of salient patterns in a time series, such as stock prices of a company over a week. model for this task should be able to extractA high-level patterns such as presence of a peak or a dip. While typical contemporary neural models with attention mechanisms can gener- ate fluent output descriptions for this task, they often generate factually incorrect descriptions. We propose a computational model with a truth-conditional architecture which first runs small learned programs on the input time series, then identifies the programs/patterns which hold true for the given input, and fi- nally conditions on only the chosen valid pro- gram (rather than the input time series) to gen- erate the output text description. A program in our model is constructed from modules, which are small neural networks that are designed to capture numerical patterns and temporal infor- mation. The modules are shared across mul- tiple programs, enabling compositionality as well as efficient learning of module parame- ters. The modules, as well as the composition of the modules, are unobserved in data, and we learn them in an end-to-end fashion with the only training signal coming from the accompa- nying natural language text descriptions. We find that the proposed model is able to gen- erate high-precision captions even though we consider a small and simple space of module types. 1 Introduction There has been large interest in generating auto- matic text description (McKeown, 1992) of tabu- lar data – for example, prior work has sought to generate biographies from tables of biographical information (Lebret et al., 2016), and generating descriptions from structured meaning representa- tions (Clairet, 2017). However, in many of these tasks the main focus is on designing systems that Figure 1: We propose a neural truth conditional model for high precision and diverse time series caption generation. are able to select entries from tabular or equivalent data during generation by using neural attention mechanisms. In many naturally occurring descrip- tions of tabular data, humans often refer to higher- level patterns, for example in the description of stock index pricing over the week in Fig. 1, the speaker refers to how the stock price peaks towards the ending. Some recent work has looked into se- tups which require non-trivial inference (Wiseman et al., 2017; Chen et al., 2020). However, they typically don’t involve inference about numerical patterns in time series data. Moreover, much recent prior work on identifying more complex patterns in data for captioning has relied on deep neural networks, often employing neural encoders and at- tention mechanisms. However, such approaches often fail to generate faithful responses and lack interpretability (Tian …

Content truncated to 500 words