When FLUE Meets FLANG: Benchmarks and Large Pretrained Language Model for Financial Domain

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 184 - 198 December 7-11, 2022 ©2022 Association for Computational Linguistics Interpreting Language Models with Contrastive Explanations Kayo Yin∗ University of California, Berkeley kayoyin@berkeley.edu Graham Neubig Carnegie Mellon University gneubig@cs.cmu.edu Abstract Model interpretability methods are often used to explain NLP model decisions on tasks such as text classification, where the output space is relatively small. However, when applied to language generation, where the output space often consists of tens of thousands of tokens, these methods are unable to provide informa- tive explanations. Language models must con- sider various features to predict a token, such as its part of speech, number, tense, or seman- tics. Existing explanation methods conflate ev- idence for all these features into a single expla- nation, which is less interpretable for human understanding. To disentangle the different decisions in lan- guage modeling, we focus on explaining lan- guage models contrastively: we look for salient input tokens that explain why the model predicted one token instead of another. We demonstrate that contrastive explanations are quantifiably better than non-contrastive expla- nations in verifying major grammatical phe- nomena, and that they significantly improve contrastive model simulatability for human ob- servers. We also identify groups of contrastive decisions where the model uses similar evi- dence, and we are able to characterize what in- put tokens models use during various language generation decisions.1 1 Introduction Despite their success across a wide swath of natural language processing (NLP) tasks, neural language models (LMs) are often used as black boxes: how they make certain predictions remains obscure (Be- linkov and Glass, 2019). This is in part due to the high complexity of the LM task itself, as well as that of the model architectures used to solve it. We argue that this is also due to the fact that inter- pretability methods commonly used in NLP, such ∗Work done while at Carnegie Mellon University. 1Code and demo: https://github.com/kayoyin/interpret-lm. Input: Can you stop the dog from Output: barking 1. Why did the model predict “barking”? Can you stop the dog from 2. Why did the model predict “barking” instead of “crying”? Can you stop the dog from 3. Why did the model predict “barking” instead of “walking”? Can you stop the dog from Table 1: Explanations for the GPT-2 prediction given the input “Can you stop the dog from _____". Input to- kens that are measured to raise or lower the probability of “barking” are in red and blue respectively, and those with little influence are in white. Non-contrastive ex- planations such as gradient × input (1) usually attribute the highest saliency to the token immediately preceding the prediction. Contrastive explanations (2, 3) give a more fine-grained and informative explanation on why the model predicted one token over another. as gradient-based saliency maps (Li et al., 2016a; Sundararajan et al., 2017), are not as informative for LM predictions compared to other tasks like text classification. For example, to explain why an LM …

Content truncated to 500 words