CausaLM: Causal Model Explanation Through Counterfactual Language Models
AbstractUnderstanding predictions made by deep neural networks is notoriously difficult, but also crucial to their dissemination. As all machine learning–based methods, they are as good as their training data, and can also capture unwanted biases. While there are tools that can help...
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
The MIT Press
2021-07-01
|
| Series: | Computational Linguistics |
| Online Access: | https://direct.mit.edu/coli/article/47/2/333/98518/CausaLM-Causal-Model-Explanation-Through |
| _version_ | 1811251409562107904 |
|---|---|
| author | Amir Feder Nadav Oved Uri Shalit Roi Reichart |
| author_facet | Amir Feder Nadav Oved Uri Shalit Roi Reichart |
| author_sort | Amir Feder |
| collection | DOAJ |
| description |
AbstractUnderstanding predictions made by deep neural networks is notoriously difficult, but also crucial to their dissemination. As all machine learning–based methods, they are as good as their training data, and can also capture unwanted biases. While there are tools that can help understand whether such biases exist, they do not distinguish between correlation and causation, and might be ill-suited for text-based models and for reasoning about high-level language concepts. A key problem of estimating the causal effect of a concept of interest on a given model is that this estimation requires the generation of counterfactual examples, which is challenging with existing generation technology. To bridge that gap, we propose CausaLM, a framework for producing causal model explanations using counterfactual language representation models. Our approach is based on fine-tuning of deep contextualized embedding models with auxiliary adversarial tasks derived from the causal graph of the problem. Concretely, we show that by carefully choosing auxiliary adversarial pre-training tasks, language representation models such as BERT can effectively learn a counterfactual representation for a given concept of interest, and be used to estimate its true causal effect on model performance. A byproduct of our method is a language representation model that is unaffected by the tested concept, which can be useful in mitigating unwanted bias ingrained in the data.1 |
| first_indexed | 2024-04-12T16:19:12Z |
| format | Article |
| id | doaj.art-695c189dcd6c464bb43626363309264e |
| institution | Directory Open Access Journal |
| issn | 0891-2017 1530-9312 |
| language | English |
| last_indexed | 2024-04-12T16:19:12Z |
| publishDate | 2021-07-01 |
| publisher | The MIT Press |
| record_format | Article |
| series | Computational Linguistics |
| spelling | doaj.art-695c189dcd6c464bb43626363309264e2022-12-22T03:25:37ZengThe MIT PressComputational Linguistics0891-20171530-93122021-07-0147233338610.1162/coli_a_00404CausaLM: Causal Model Explanation Through Counterfactual Language ModelsAmir Feder0Nadav Oved1Uri Shalit2Roi Reichart3Faculty of Industrial Engineering and Management, Technion - Israel Institute of Technology. feder@campus.technion.ac.ilFaculty of Industrial Engineering and Management, Technion - Israel Institute of Technology. nadavo@campus.technion.ac.ilFaculty of Industrial Engineering and Management, Technion - Israel Institute of Technology. urishalit@technion.ac.ilFaculty of Industrial Engineering and Management, Technion - Israel Institute of Technology. roiri@technion.ac.il AbstractUnderstanding predictions made by deep neural networks is notoriously difficult, but also crucial to their dissemination. As all machine learning–based methods, they are as good as their training data, and can also capture unwanted biases. While there are tools that can help understand whether such biases exist, they do not distinguish between correlation and causation, and might be ill-suited for text-based models and for reasoning about high-level language concepts. A key problem of estimating the causal effect of a concept of interest on a given model is that this estimation requires the generation of counterfactual examples, which is challenging with existing generation technology. To bridge that gap, we propose CausaLM, a framework for producing causal model explanations using counterfactual language representation models. Our approach is based on fine-tuning of deep contextualized embedding models with auxiliary adversarial tasks derived from the causal graph of the problem. Concretely, we show that by carefully choosing auxiliary adversarial pre-training tasks, language representation models such as BERT can effectively learn a counterfactual representation for a given concept of interest, and be used to estimate its true causal effect on model performance. A byproduct of our method is a language representation model that is unaffected by the tested concept, which can be useful in mitigating unwanted bias ingrained in the data.1https://direct.mit.edu/coli/article/47/2/333/98518/CausaLM-Causal-Model-Explanation-Through |
| spellingShingle | Amir Feder Nadav Oved Uri Shalit Roi Reichart CausaLM: Causal Model Explanation Through Counterfactual Language Models Computational Linguistics |
| title | CausaLM: Causal Model Explanation Through Counterfactual Language Models |
| title_full | CausaLM: Causal Model Explanation Through Counterfactual Language Models |
| title_fullStr | CausaLM: Causal Model Explanation Through Counterfactual Language Models |
| title_full_unstemmed | CausaLM: Causal Model Explanation Through Counterfactual Language Models |
| title_short | CausaLM: Causal Model Explanation Through Counterfactual Language Models |
| title_sort | causalm causal model explanation through counterfactual language models |
| url | https://direct.mit.edu/coli/article/47/2/333/98518/CausaLM-Causal-Model-Explanation-Through |
| work_keys_str_mv | AT amirfeder causalmcausalmodelexplanationthroughcounterfactuallanguagemodels AT nadavoved causalmcausalmodelexplanationthroughcounterfactuallanguagemodels AT urishalit causalmcausalmodelexplanationthroughcounterfactuallanguagemodels AT roireichart causalmcausalmodelexplanationthroughcounterfactuallanguagemodels |