Anomaly Detection with Neural Parsers That Never Reject

Abstract: Reinforcement learning has recently shown promise as a technique for training an artificial neural network to parse sentences in some unknown format, through a body of work known as RL-GRIT. A key aspect of the RL-GRIT approach is that rather than explicitly inferring a grammar that describes the format, the neural network learns to perform various parsing actions (such as merging two tokens) over a corpus of sentences, with the goal of maximizing the estimated frequency of the resulting parse structures. This can allow the learning process to more easily explore different action choices, since a given choice may change the optimality of the parse (as expressed by the total reward), but will not result in the failure to parse a sentence. However, this also presents a limitation: because the trained neural network can successfully parse any sentence, it cannot be directly used to identify sentences that deviate from the format of the training sentences, i.e., that are anomalous. In this paper, we address this limitation by presenting procedures for extracting production rules from the neural network, and for using these rules to determine whether a given sentence is nominal or anomalous. When a sentence is anomalous, an attempt is made to identify the location of the anomaly. We empirically demonstrate that our approach is capable of grammatical inference and anomaly detection for both non-regular formats and those containing regions of high randomness/entropy. While a format with high randomness typically requires large sets of production rules, we propose a two pass grammatical inference method to generate parsimonious rule sets for such formats. By further improving parser learning, and leveraging the presented rule extraction and anomaly detection algorithms, one might begin to understand common errors, either benign or malicious, in practical formats.