HILDIF: Interactive Debugging of NLI Models Using Influence Functions

Abstract

Biases and artifacts in training data can cause unwelcome behavior in text classifiers (such as shallow pattern matching), leading to lack of generalizability. One solution to this problem is to include users in the loop and leverage their feedback to improve models. We propose a novel explanatory debugging pipeline called HILDIF, enabling humans to improve deep text classifiers using influence functions as an explanation method. We experiment on the Natural Language Inference (NLI) task, showing that HILDIF can effectively alleviate artifact problems in fine-tuned BERT models and result in increased model generalizability.

Explanation-Based Human Debugging of NLP Models: A Survey

“…Most papers in Table 1 focus on text classification with single input (TC) for a variety of specific problems such as email categorization , topic classification (Kulesza et al, 2015;Teso and Kersting, 2019), spam classification (Koh and Liang, 2017), sentiment analysis (Ribeiro et al, 2018b), and auto-coding of transcripts (Kulesza et al, 2010). By contrast, Zylberajch et al (2021) Ghai et al (2021) suggested that most researchers work on TC because, for this task, it is much easier for lay participants to understand explanations and give feedback (e.g., which keywords should be added or removed from the list of top features). 4 Meanwhile, some other NLP tasks require the feedback providers to have linguistic knowledge such as part-of-speech tagging, parsing, and machine translation.…”

“…Some papers even allowed humans to adjust the word importance scores (WS) (Kulesza et al, 2009(Kulesza et al, , 2015. This is analogous to specifying relevancy scores for example-based explanations (ES) in Zylberajch et al (2021). Meanwhile, feedback at the level of learned features (FE) (i.e., the internal neurons in the model) and learned rules (RU) rather than individual words, was asked in Lertvittayakumjorn et al (2020) andRibeiro et al (2018b), respectively.…”

“…To conduct experiments, some studies in Table 1 selected human participants (SP) to be their feedback providers. The selected participants could be people without ML/NLP knowledge (Kulesza et al, 2010(Kulesza et al, , 2015 or with ML/NLP knowledge (Ribeiro et al, 2018b;Zylberajch et al, 2021) depending on the study objectives and the complexity of the feedback process. Early work even conducted experiments with the participants inperson Kulesza et al, 2009Kulesza et al, , 2015.…”

“…However, we suggest that there are several tasks where the trained models are prone to be buggy but the tasks are underexplored in the EBHD setting, though they are not too difficult to experiment on with lay people. NLI, the focus of Zylberajch et al (2021), is one of them. Indeed, McCoy et al (2019) and Gururangan et al (2018) showed that NLI models can exploit annotation artifacts and fallible syntactic heuristics to make predictions rather than learning the logic of the actual task.…”

“…pers (ranked by relevancy) for each query and kept only the ones appearing in at least 2 out of the 16 query results. This resulted in 234 papers that we then manually checked, leading to selecting a few additional papers, including Han and Ghosh (2020) and Zylberajch et al (2021). The overall process resulted in 15 papers listed in Table 1 as the selected studies primarily discussed in this survey.…”