Generating and Modifying Natural Language Explanations

Generating and Modifying Natural Language Explanations Abdus Salam and Rolf Schwitter and Mehmet A. Orgun Macquarie University, Sydney, Australia {abdus.salam, rolf.schwitter, mehmet.orgun}@mq.edu.au Abstract HESIP is a hybrid explanation system for im- age predictions that combines sub-symbolic and symbolic machine learning techniques to explain the predictions of image classification tasks. The sub-symbolic component makes a prediction for an image and the symbolic com- ponent learns probabilistic symbolic rules in order to explain that prediction. In HESIP, the explanations are generated in controlled natural language from the learned probabilistic rules using a bi-directional logic grammar. In this paper, we present an explanation modification method where a human-in-the-loop can mod- ify an incorrect explanation generated by the HESIP system and afterwards, the modified ex- planation is used by the symbolic component of HESIP to learn a better explanation. 1 Introduction In recent years, the development of explanation systems has gained a lot of attention. Most of these explanation systems (Ribeiro et al., 2016, 2018; Lundberg and Lee, 2017) can explain pre- dictions made by machine learning (ML) models. Researchers are focusing on building explanation systems for ML models, because these models have shown excellent performance for different predic- tion tasks (Zhang et al., 2020; LeCun et al., 2015) and most of these models are sub-symbolic black- box models that are not easily understandable, and therefore lead to difficulties explaining the predic- tions to the users. Explanation systems such as Lime (Ribeiro et al., 2016), Anchor (Ribeiro et al., 2018) and SHAP (Lundberg and Lee, 2017) use existing information of the datasets to explain pre- dictions. However, sometime information that is not present directly in the dataset such as relation information can play an important role in the ex- planation; especially, in image prediction tasks as shown in LIME-Aleph (Rabold et al., 2019). HESIP is a hybrid explanation system for im- age predictions. The HESIP system explains the predictions to the users using natural language ex- planations. The explanations are generated in a con- trolled natural language (CNL) (Kuhn, 2014) using a logic programming based bi-directional gram- mar that is similar to Schwitter (2018). The gener- ated explanations of the HESIP system are human- understandable as well as machine-processable. Since the explanations are represented in a natural language, they are immediately understandable by all types of users. The bi-directional grammar of the HESIP system can also process a generated ex- planation that has been modified by the user. The HESIP system aims to generate an explanation for the predicted image that represents the object in- formation together with the relation information. It is expected that such as system is not perfect, and HESIP is not an exception. HESIP sometimes generates wrong explanations. To the best of our knowledge, there is no explanation system that al- lows a user to modify an explanation in order to improve the explanation generation process of the system. It is important that a user can modify an incorrect explanation so that the system can learn how to generate a …

Content truncated to 500 words