Visualizing the Relationship Between Encoded Linguistic Information and Task Performance

Findings of the Association for Computational Linguistics: ACL 2022, pages 410 - 422 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Visualizing the Relationship Between Encoded Linguistic Information and Task Performance Jiannan Xiang♥∗, Huayang Li♠∗, Defu Lian♦, Guoping Huang♣, Taro Watanabe♠, Lemao Liu♣ ♥Carnegie Mellon University ♠Nara Institute of Science and Technology ♦University of Science and Techology of China ♣Tencent AI Lab jiannanx@cs.cmu.edu, li.huayang.lh6@is.naist.jp liandefu@ustc.edu.cn, donkeyhuang@tencent.com taro@is.naist.jp, lemaoliu@gmail.com Abstract Probing is popular to analyze whether linguis- tic information can be captured by a well- trained deep neural model, but it is hard to an- swer how the change of the encoded linguis- tic information will affect task performance. To this end, we study the dynamic relation- ship between the encoded linguistic informa- tion and task performance from the viewpoint of Pareto Optimality. Its key idea is to ob- tain a set of models which are Pareto-optimal in terms of both objectives. From this view- point, we propose a method to optimize the Pareto-optimal models by formalizing it as a multi-objective optimization problem. We con- duct experiments on two popular NLP tasks, i.e., machine translation and language model- ing, and investigate the relationship between several kinds of linguistic information and task performances. Experimental results demon- strate that the proposed method is better than a baseline method. Our empirical findings sug- gest that some syntactic information is helpful for NLP tasks whereas encoding more syntac- tic information does not necessarily lead to bet- ter performance, because the model architec- ture is also an important factor. 1 Introduction Recent years have witnessed great success of deep neural networks for natural language process- ing tasks, such as language modeling (Zaremba et al., 2014; Merity et al., 2018) and Neural Ma- chine Translation (Bahdanau et al., 2015; Vaswani et al., 2017). The excellent task performance they achieved spiked the interest in interpreting their un- derlying mechanism. Since linguistic knowledge is crucial in natural languages, an emerging body of literature uses probes (Conneau et al., 2018; Alt et al., 2020; Saleh et al., 2020; Cao et al., 2021) to investigate whether a standard model trained ∗Equal contribution. Work done while J. Xiang was an intern at Tencent AI Lab. <latexit sha1_base64="jxqvBMJlwqJp5JVo5sgYCLxCmck=">AB83icbVDLSgMxFL3js9ZX1aWbYBHcWGZU1GXRjQsXFew DOkPJpJk2NMkMSUYoQ3/DjQtF3Poz7vwbM+0stPVA4HDOvdyTEyacaeO6387S8srq2npo7y5tb2zW9nb+k4VYQ2Scxj1QmxpxJ2jTMcNpJFMUi5LQdjm5zv/1ElWaxfDTjhAYCDySLGMHGSv6pL7AZEsyz+0mvUnVr7hRokXgFqUKBRq/y5fdjkgoqDeFY67n JibIsDKMcDop+6mCSYjPKBdSyUWVAfZNPMEHVulj6JY2ScNmq/NzIstB6L0E7mEfW8l4v/ed3URNdBxmSGirJ7FCUcmRilBeA+kxRYvjYEkwUs1kRGWKFibE1lW0J3vyXF0nrOZd1s4fLqr1m6KOEhzCEZyAB1dQhztoQBMIJPAMr/DmpM6L8+58zEaXnGLnA P7A+fwB7ryRoQ=</latexit>−L <latexit sha1_base64="A4PIrUM/ilCM4eCLzZpiYP+Y618=">AB8nicbVDLSgMxFM3UV62vqks3wSK4KjMq 6rLoRncV7AOmQ8mkmTY0kwzJHaEM/Qw3LhRx69e482/MtLPQ1gOBwzn3knNPmAhuwHW/ndLK6tr6RnmzsrW9s7tX3T9oG5VqylpUCaW7ITFMcMlawEGwbqIZiUPBOuH4Nvc7T0wbruQjTBIWxGQoecQpASv5vZjAiBKR3U/71Zpbd2fA y8QrSA0VaParX72BomnMJFBjPE9N4EgIxo4FWxa6aWGJYSOyZD5lkoSMxNks8hTfGKVAY6Utk8Cnqm/NzISGzOJQzuZRzSLXi7+5/kpRNdBxmWSApN0/lGUCgwK5/fjAdeMgphYQqjmNiumI6IJBdtSxZbgLZ68TNpnde+yfv5wUWvc FHWU0RE6RqfIQ1eoge5QE7UQRQo9o1f05oDz4rw7H/PRklPsHKI/cD5/AH8SkWc=</latexit>I Figure 1: Illustration of Pareto frontier by a toy exam- ple. Triangle (△) corresponds to the standard check- point with best performance and each circle (⃝) corre- sponds to a sampled checkpoint. The y-axis indicates the linguistic information I encoded by the model, and x-axis indicates the negative loss value −L. towards better task performance also captures the linguistic information. From the perspective of information theory, Voita and Titov (2020) and Pi- mentel et al. (2020b) show that probes can be used to estimate the amount of linguistic information captured by a fixed model. However, the above probing only extracts lin- guistic information from a fixed standard model, which helps little to understand the relationship between the task performance and linguistic infor- mation encoded by the model. …

Content truncated to 500 words