Causal Direction of Data Collection Matters: Implications of Causal and Anticausal Learning for NLP

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9499–9513 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 9499 Causal Direction of Data Collection Matters: Implications of Causal and Anticausal Learning for NLP Zhijing Jin1,2∗, Julius von Kügelgen1,3∗, Jingwei Ni4, Tejas Vaidhya5, Ayush Kaushal5, Mrinmaya Sachan2 and Bernhard Schölkopf1,2 1Max Planck Institute for Intelligent Systems, Tübingen, Germany, 2ETH Zürich, 3University of Cambridge, 4University College London, 5IIT Kharagpur {zjin,jvk,bs}@tue.mpg.de, ucabjni@ucl.ac.uk, {ayushkaushal, iamtejasvaidhya}@iitkgp.ac.in, msachan@ethz.ch Abstract The principle of independent causal mecha- nisms (ICM) states that generative processes of real world data consist of independent mod- ules which do not influence or inform each other. While this idea has led to fruitful de- velopments in the field of causal inference, it is not widely-known in the NLP community. In this work, we argue that the causal direc- tion of the data collection process bears non- trivial implications that can explain a number of published NLP findings, such as differences in semi-supervised learning (SSL) and domain adaptation (DA) performance across different settings. We categorize common NLP tasks according to their causal direction and empir- ically assay the validity of the ICM princi- ple for text data using minimum description length. We conduct an extensive meta-analysis of over 100 published SSL and 30 DA stud- ies, and find that the results are consistent with our expectations based on causal insights. This work presents the first attempt to analyze the ICM principle in NLP, and provides construc- tive suggestions for future modeling choices.1 1 Introduction NLP practitioners typically do not pay great atten- tion to the causal direction of the data collection process. As a motivating example, consider the case of collecting a dataset to train a machine trans- lation (MT) model to translate from English (En) to Spanish (Es): it is common practice to mix all avail- able En-Es sentence pairs together and train the model on the entire pooled data set (Bahdanau et al., 2015; Cho et al., 2014). However, such mixed cor- pora actually consist of two distinct types of data: (i) sentences that originated in English and have been translated (by human translators) into Span- ish (En→Es); and (ii) sentences that originated in ∗Equal contribution. 1The codes are at https://github.com/zhijing-jin/icm4nlp. Given the English sentence above, can you write its Spanish translation? Prompt for annotators [En] This is a beautiful world. [Es] Este es un mundo hermoso. Cause: Effect: Annotation process (Noise) Effect = CausalMechanism (Cause, Noise) Figure 1: Annotation process for NLP data: the ran- dom variable that exists first is typically the cause (e.g., a given prompt), and the one generated afterwards is typically the effect (e.g., the annotated answer). Spanish and have subsequently been translated into English (Es→En).2 Intuitively, these two subsets are qualitatively dif- ferent, and an increasing number of observations by the NLP community indeed suggests that they exhibit different properties (Freitag et al., 2019; Edunov et al., 2020; Riley et al., 2020; Shen et al., 2021). In the case of MT, for example, …

Content truncated to 500 words