Context Dependent Semantic Parsing over Temporally Structured Data

Proceedings of NAACL-HLT 2019, pages 3576–3585 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3576 Context-Dependent Semantic Parsing over Temporally Structured Data Charles Chen and Razvan Bunescu School of Electrical Engineering and Computer Science, Ohio University lc971015@ohio.edu, bunescu@ohio.edu Abstract We describe a new semantic parsing setting that allows users to query the system using both natural language questions and actions within a graphical user interface. Multiple time series belonging to an entity of interest are stored in a database and the user interacts with the system to obtain a better understanding of the entity’s state and behavior, entailing se- quences of actions and questions whose an- swers may depend on previous factual or nav- igational interactions. We design an LSTM- based encoder-decoder architecture that mod- els context dependency through copying mech- anisms and multiple levels of attention over inputs and previous outputs. When trained to predict tokens using supervised learning, the proposed architecture substantially outper- forms standard sequence generation baselines. Training the architecture using policy gradient leads to further improvements in performance, reaching a sequence-level accuracy of 88.7% on artificial data and 74.8% on real data. 1 Introduction and Motivation Wearable sensors are being increasingly used in medicine to monitor important physiological pa- rameters. Patients with type I diabetes, for exam- ple, wear a sensor inserted under the skin which provides measurements of the interstitial blood glu- cose level (BGL) every 5 minutes. Sensor bands provide a non-invasive solution to measuring addi- tional physiological parameters, such as tempera- ture, skin conductivity, heart rate, and acceleration of body movements. Patients may also self-report information about discrete life events such as meals, sleep, or stressful events, while an insulin pump automatically records two types of insulin inter- ventions: a continuous stream of insulin called the basal rate, and discrete self-administered insulin dosages called boluses. The data acquired from sen- sors and patients accumulates rapidly and leads to a substantial data overload for the health provider. To help doctors more easily browse the wealth of generated patient data, we built a graphical user in- terface (GUI) that displays the various time series of measurements corresponding to a patient. As shown in Figure 1, the GUI displays the data cor- responding to one day, whereas buttons allow the user to move to the next or previous day. While the graphical interface was enthusiastically received by doctors, it soon became apparent that the doctor- GUI interaction could be improved substantially if the tool also allowed for natural language (NL) interactions. Most information needs are highly contextual and local. For example, if the blood glu- cose spiked after a meal, the doctor would often want to know more details about the meal or about the bolus that preceded the meal. The doctor often found it easier to express their queries in natural language (e.g. “show me how much he ate", “did he bolus before that"), resulting in a sub-optimal situ- ation where the doctor would ask this type of local questions …

Content truncated to 500 words