Explaining non-linear Classifier Decisions within Kernel-based Deep Architectures

Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 16–24 Brussels, Belgium, November 1, 2018. c⃝2018 Association for Computational Linguistics 16 Explaining non-linear Classifier Decisions within Kernel-based Deep Architectures Danilo Croce and Daniele Rossini and Roberto Basili Department of Enterprise Engineering University of Roma, Tor Vergata {croce,basili}@info.uniroma2.it Abstract Nonlinear methods such as deep neural net- works achieve state-of-the-art performances in several semantic NLP tasks. However episte- mologically transparent decisions are not pro- vided as for the limited interpretability of the underlying acquired neural models. In neural- based semantic inference tasks epistemologi- cal transparency corresponds to the ability of tracing back causal connections between the linguistic properties of a input instance and the produced classification output. In this paper, we propose the use of a method- ology, called Layerwise Relevance Propaga- tion, over linguistically motivated neural ar- chitectures, namely Kernel-based Deep Archi- tectures (KDA), to guide argumentations and explanation inferences. In such a way, each decision provided by a KDA can be linked to real examples, linguistically related to the in- put instance: these can be used to motivate the network output. Quantitative analysis shows that richer explanations about the semantic and syntagmatic structures of the examples charac- terize more convincing arguments in two tasks, i.e. question classification and semantic role labeling. 1 Introduction Nonlinear methods such as deep neural networks achieve state-of-the-art performances in several challenging problems, such as image classification or natural language processing (NLP). However the traditional AI criticism still holds: they are not epistemologically transparent, as for the limited interpretability of the neural inferences. In a question classification (QC) task, e.g. (Li and Roth, 2006), this is particularly evident. The category describing the target of a request is rel- evant in question answering to optimize the later stages of search and answer detection, and its in- terpretation depends on a variety of semantic and syntactic properties of the question. Epistemolog- ical transparency corresponds here to the ability of tracing back the connections between linguistic properties of the input question and the proposed question category. An example-driven machine learning model should be able to provide causal re- lations between the input semantic aspect and the properties of the question. For example, given the prediction ”What is the capital of Zimbabwe?” refers to a Location, we would like the system to motivate it with a sen- tence such as: Since it seems similar to ”What is the capital of California?” which also refers to a Location. Notice how in neural learning, as for exam- ple in Multilayer Perceptrons, Long Short-Term Memory Networks, (Hochreiter and Schmidhuber, 1997), or the more recent Attention-based Net- works (Larochelle and Hinton, 2010), the network parameters have no clear conceptual counterpart. Using the Layerwise Relevance Propagation (LRP) (Bach et al., 2015) approach, the classi- fication decisions of a multilayer perceptron are decomposed backward across the network layers, and evidence about the contribution of individual input fragments (i.e. layer 0) to the final decision is gathered. Evaluation against images (i.e. …

Content truncated to 500 words