Jim Hunter

Choosing words in computer-generated weather forecasts

One of the main challenges in automatically generating textual weather forecasts is choosing appropriate English words to communicate numeric weather data. A corpus-based analysis of how humans write forecasts showed that there were major differences in how individual writers performed this task, that is, in how they translated data into words. These differences included both different preferences between potential near-synonyms that could be used to express information, and also differences in the meanings that individual writers associated with specific words. Because we thought these differences could confuse readers, we built our SumTime-Mousam weather-forecast generator to use consistent data-to-word rules, which avoided words which were only used by a few people, and words which were interpreted differently by different people. An evaluation by forecast users suggested that they preferred SumTime-Mousams texts to human-generated texts, in part because of better word choice; this may be the first time that an evaluation has shown that nlg texts are better than human-authored texts.

Choosing the content of textual summaries of large time-series data sets

Natural Language Generation (NLG) can be used to generate textual summaries of numeric data sets. In this paper we develop an architecture for generating short (a few sentences) summaries of large (100KB or more) time-series data sets. The architecture integrates pattern recognition, pattern abstraction, selection of the most significant patterns, microplanning (especially aggregation), and realisation. We also describe and evaluate SumTime-Turbine, a prototype system which uses this architecture to generate textualsummaries of sensor data from gas turbines.

Knowledge-Based Event Detection in Complex Time Series Data

This paper describes an approach to the detection of events in complex, multi-channel, high frequency data. The example used is that of detecting the re-siting of a transcutaneous O2/CO2 probe on a baby in a neonatal intensive care unit (ICU) from the available monitor data. A software workbench has been developed which enables the expert clinician to display the data and to mark up features of interest. This knowledge is then used to define the parameters for a pattern matcher which runs over a set of intervals derived from the raw data by a new iterative interval merging algorithm. The approach has been tested on a set of 45 probe changes; the preliminary results are encouraging, with an accuracy of identification of 89%.

From data to text in the Neonatal Intensive Care Unit: Using NLG technology for decision support and information management

Contemporary Neonatal Intensive Care Units collect vast amounts of patient data in various formats, making efficient processing of information by medical professionals difficult. Moreover, different stakeholders in the neonatal scenario, which include parents as well as staff occupying different roles, have different information requirements. This paper describes recent and ongoing work on building systems that automatically generate textual summaries of neonatal data. Our evaluation results show that the technology is viable and comparable in its effectiveness for decision support to existing presentation modalities. We discuss the lessons learned so far, as well as the major challenges involved in extending current technology to deal with a broader range of data types, and to improve the textual output in the form of more coherent summaries.

A comparison of graphical and textual presentations of time series data to support medical decision making in the neonatal intensive care unit.

Objective. To compare expert-generated textual summaries of physiological data with trend graphs, in terms of their ability to support neonatal Intensive Care Unit (ICU) staff in making decisions when presented with medical scenarios. Methods. Forty neonatal ICU staff were recruited for the experiment, eight from each of five groups – junior, intermediate and senior nurses, junior and senior doctors. The participants were presented with medical scenarios on a computer screen, and asked to choose from a list of 18 possible actions those they thought were appropriate. Half of the scenarios were presented as trend graphs, while the other half were presented as passages of text. The textual summaries had been generated by two human experts and were intended to describe the physiological state of the patient over a short period of time (around 40 minutes) but not to interpret it. Results. In terms of the content of responses there was a clear advantage for the Text condition, with participants tending to choose more of the appropriate actions when the information was presented as text rather than as graphs. In terms of the speed of response there was no difference between the Graphs and Text conditions. There was no significant difference between the staff groups in terms of speed or content of responses. In contrast to the objective measures of performance, the majority of participants reported a subjective preference for the Graphs condition. Conclusions. In this experimental task, participants performed better when presented with a textual summary of the medical scenario than when it was presented as a set of trend graphs. If the necessary algorithms could be developed that would allow computers automatically to generate descriptive summaries of physiological data, this could potentially be a useful feature of decision support tools in the intensive care unit.

Generating English summaries of time series data using the Gricean maxims

We are developing technology for generating English textual summaries of time-series data, in three domains: weather forecasts, gas-turbine sensor readings, and hospital intensive care data. Our weather-forecast generator is currently operational and being used daily by a meteorological company. We generate summaries in three steps: (a) selecting the most important trends and patterns to communicate; (b) mapping these patterns onto words and phrases; and (c) generating actual texts based on these words and phrases. In this paper we focus on the first step, (a), selecting the information to communicate, and describe how we perform this using modified versions of standard data analysis algorithms such as segmentation. The modifications arose out of empirical work with users and domain experts, and in fact can all be regarded as applications of the Gricean maxims of Quality, Quantity, Relevance, and Manner, which describe how a cooperative speaker should behave in order to help a hearer correctly interpret a text. The Gricean maxims are perhaps a key element of adapting data analysis algorithms for effective communication of information to human users, and should be considered by other researchers interested in communicating data to human users.

Deriving Trends in Historical and Real-Time Continuously Sampled Medical Data

Monitors in Intensive Care Units generate large volumes of continuous data which can overwhelm a database and result in information overload for the medical staff. Instead of reasoning with individual data samples of one or more variables, it is better to work with the trend of the data i.e., whether the data is increasing, decreasing or steady. We have developed a system which abstracts continuous data into trends; it consists of three consecutive processes: filtering which smooths the data; temporal interpolation which creates simple intervals between consecutive data points; and temporal inference which iteratively merges intervals which share similar characteristics into larger intervals. Storing trends can result in a reduction in database volume. Our system has been applied both to historical and real-time data.

Automatic generation of natural language nursing shift summaries in neonatal intensive care: BT-Nurse

INTRODUCTION Our objective was to determine whether and how a computer system could automatically generate helpful natural language nursing shift summaries solely from an electronic patient record system, in a neonatal intensive care unit (NICU). METHODS A system was developed which automatically generates partial NICU shift summaries (for the respiratory and cardiovascular systems), using data-to-text technology. It was evaluated for 2 months in the NICU at the Royal Infirmary of Edinburgh, under supervision. RESULTS In an on-ward evaluation, a substantial majority of the summaries was found by outgoing and incoming nurses to be understandable (90%), and a majority was found to be accurate (70%), and helpful (59%). The evaluation also served to identify some outstanding issues, especially with regard to extra content the nurses wanted to see in the computer-generated summaries. CONCLUSIONS It is technically possible automatically to generate limited natural language NICU shift summaries from an electronic patient record. However, it proved difficult to handle electronic data that was intended primarily for display to the medical staff, and considerable engineering effort would be required to create a deployable system from our proof-of-concept software.

Automatic Generation of Textual Summaries from Neonatal Intensive Care Data

Intensive care is becoming increasingly complex. If mistakes are to be avoided, there is a need for the large amount of clinical data to be presented effectively to the medical staff. Although the most common approach is to present the data graphically, it has been shown that textual summarisation can lead to improved decision making. As the first step in the BabyTalk project, a prototype is being developed which will generate a textual summary of 45 minutes of continuous physiological signals and discrete events (e.g.: equipment settings and drug administration). Its architecture brings together techniques from the different areas of signal analysis, medical reasoning, and natural language generation. Although the current system is still being improved, it is powerful enough to generate meaningful texts containing the most relevant information. This prototype will be extended to summarize several hours of data and to include clinical interpretation.

Segmenting Time Series for Weather Forecasting

We are investigating techniques for producing textual summaries of time series data. Deep reasoning techniques have proven impractical because we lack perfect knowledge about users and their tasks. Data analysis techniques such as segmentation are more attractive, but they have been developed for data mining, not for communication. We examine how segmentation should be modified to make it suitable for generating textual summaries. Our algorithm has been implemented in a weather forecast generation system.