OntoED: Low-resource Event Detection with Ontology Embedding

To categorize and synthesize medication safety event detection methods in the critically ill in order to provide clinicians and administrators with approaches to event detection that are intended to expand and complement traditional voluntary reporting systems. A literature search of OvidMEDLINE was performed to identify articles related to medication safety involving critically ill patients in the intensive care unit setting. The inclusion of articles was restricted to comparative studies. The bibliographies of all retrieved articles were reviewed to obtain additional articles of relevance. The various event detection methods were compared by: evidence supporting their use; number, type and severity of events detected; phase of the medication use process in which events were detected; and ease and cost of implementation. Major limitations of each method were also collated. There are a number of methods that can be used to identify medication safety events in the critically ill. These can broadly be categorized as: 1) voluntary reporting, 2) record review, 3) rules/triggers and 4) direct observation and 5) interviews/surveys. Relatively few studies have directly compared these assessment methods in the ICU setting, although the limitations of the traditional voluntary reporting system as the sole method of event detection are well established. Although not truly dichotomous, these methods can be broken down into more proactive and reactive approaches. Rules/triggers and direct observation of the medication use process in the ICU are examples of proactive approaches to event detection, while the traditional unsolicited voluntary reporting is typically reactive. However, each of the event detection methods has advantages and disadvantages, so the methods should not be considered mutually exclusive with respect to obtaining information about medication safety. Given the limitations of traditional voluntary reporting systems, a multimodal approach used to identify medication safety events is most likely to capture the largest number and type of events. We would advise not trying to implement additional approaches beyond voluntary reporting systems all at once. This would be difficult and costly. Rather, we suggest a systematic implementation of additional event detection approaches that takes into account hospital-specific considerations.

In this paper, we investigate a semi-supervised method, that is, self-training algorithm in event detection and type recognition. First, we get two kinds of feature vectors of sentences for event detection and type recognition task respectively. Second, we use self-training algorithm to filter the non-event instances and identify type for every event instance. The experimental result on the ACE2005 corpus shows that our model using self-training algorithm can identify the majority of non-event instances and also has a good precision and recall rate in event type recognition.KeywordsEvent ExtractionSemi-supervised LearningFeature Selection

We present an online multilingual system for event detection and comprehension from media feeds. The system retrieves information from news sites, aggregates them into events (event detection), and summarizes them by extracting semantic labels of its most relevant entities (event representation) in order to answer the journalism Ws: who, what, when and where. The generated events populate VLX-Stories -an event ontology- transforming unstructured text data to a structured knowledge base representation. Our system exploits an external entity Knowledge Graph (VKG) to help populate VLX-Stories. At the same time, this external knowledge graph can also be extended with a Dynamic Entity Linking (DEL) module, which detects emerging entities (EE) on unstructured data. The system is currently deployed in production and used by media producers in the editorial process, providing real-time access to breaking news. Each month, VLX-Stories detects over 9000 events from over 4000 news feeds from seven different countries and in three different languages. At the same time, it detects over 1300 EE per month, which populate VKG.

OBJECTIVES (1) To examine effects of secondary task demands on event detection and response during driving. (2) To examine the validity of surrogate methods which might be used in a laboratory environment to evaluate these effects early in the development process for advanced information systems. (3) To identify issues which limit and/or require careful interpretation in the application of event detection results to product considerations. (4) To identify issues which are at the forefront of scientific knowledge in this area, and in need of further research. METHODS Analytic methods were used to compare data derived from prior studies—and/or to make comparisons across studies. Linear mixed models, correlation, and other statistical methods were used, as appropriate. Results were drawn from several sources, and hence reflected multiple methodologies. One of the main sources of data for this work was the CAMP Driver Workload Metrics project (2006) (which was a collaborative project between GM, Ford, Toyota, Nissan, and USDOT). Three types of methods were compared: on-road, test track, and laboratory methods. In the onroad and test-track methods, 108 participants and 69 participants (respectively) drove an instrumented vehicle that was the middle car in a platoon of three vehicles. (That is, the test participants followed a lead vehicle—and also had a vehicle following them). Drivers engaged in 22 different secondary tasks while driving on public roads (or a test track), and while also responding to three different types of visual events that could potentially occur during a task. These event types included: the illumination of a CHMSL-like light on the lead vehicle, a deceleration of the lead vehicle (without brake lights), and a turn signal illumination on the follow-vehicle. In the laboratory, methods included a specially-developed variation of the Sternberg task, and two Peripheral Detection Tasks. Data were collected from 57 participants. Data from these methods are reported, compared with each other, and compared with data from other sources, where methods permitted that. Dependent variables were primarily “percent missed events”; key independent variables included task type and workload level. RESULTS Data showed that visual-manual and auditory-vocal tasks differed in their effects on event detection during driving, with visual-manual tasks (as a class) interfering to a greater extent with event detection. The effects of task demands on event detection could not, however, be fully accounted for by glance-related measurements, suggesting that there are other variables involved, such as an attentional component to event detection. Furthermore, the modifiedSternberg method emerged as a laboratory surrogate with promising predictive validity (as well as repeatability) for evaluating a task’s effects on event detection/response during driving. However, several issues emerged related to how the results of this assessment method can be interpreted and applied. These issues relate to practical matters (of task length, and task type), and also to theoretical matters (e.g., regarding how to change task design to mitigate interference with event detection). CONCLUSIONS Drivers’ responses to events while driving (particularly unexpected events) are an important element of driving performance. Early in the development of advanced information systems, it would be desirable to have a valid way to evaluate how task demands may affect event detection (so that system task designs could be modified and improved, if necessary). However, it is particularly difficult to study event detection experimentally. Nearly all existing methods of evaluating event detection and response (other than naturalistic driving studies) must be classified as surrogate methods. And most of the methods examined here used techniques that did not really qualify as presenting “completely unexpected” events. Nonetheless, with this caveat, the method which offered the most promise was that which was sensitive both to a task’s loading on visual processing and also its effects on central attention (e.g., to a cognitive/memory load). That method was the modified-Sternberg method. Issues related to interpretation and application suggested that discriminating high from low workload tasks may be a problem when Task Types are separated. However, conclusions about grouping Task Types together are offered. Further, the findings had implications for strategies that might be used during system development to improve event detection performance for a task. Additional research is needed, not only to verify the kinds of strategies for task development that would in fact lead to improved event detection performance, but also to examine issues surrounding development of methods for presenting unexpected events, and to establish a deeper understanding of processes that underlie event detection and response, especially in naturalistic driving REFERENCES Angell, L. S., Auflick, L. L., Austria, P. A., Kochhar, D. S. & Tijerina, L. (2006). “Driver workload metrics project: Task 2 final report,” submitted to U.S. Dept. of Transportation, FHA, & NHTSA under Cooperative Agreement Number DTFH61-01-X-00014.

Event detection (ED) is a critical task in information extraction, aiming to identify triggers and types. Current research has increasingly focused on fine-grained types, where a single sentence may contain multiple trigger words and event types. Previous models only considered sentence-level features, neglecting word-word features and word positional information in the text. We propose a novel labeling scheme that treats event detection as a word-word relation recognition task. In this approach, we first identify relationships between word pairs and then utilize these relationships to perform trigger and event-type detection. By adopting this method, we can efficiently and concurrently identify triggers and types in a sentence. Leveraging word pair relationships effectively addresses scenarios in which multiple trigger words appear in a single sentence. The results demonstrate that our approach outperforms several baseline models.

Event detection identifies interesting events from web pages and in this paper, a new approach is proposed to identify the event instances associated with an interested event type. The terms that are related to criminal activities, its co-occurrence terms and the associated sentences are considered from web documents. These sentence patterns are processed by POS tagging. Since, there is no knowledge on the sentences for the first instances, they are clustered using decision tree. Rules are formulated using pattern clusters. Priorities are assigned to the clusters based on the importance of patterns. The importance of the patterns defines the semantic relation towards event instances. Considering the priorities, weights are assigned for the rules. Artificial Neural Network (ANN) is used to classify the sentences to detect event instances based on the gained knowledge. Here ANN is used for training the weighted sentence patterns to learn the event instances of the specific event type. It is observed that the constructed rule is effective in classifying the sentences to identify event instance. The combination of these sentence patterns of the event instances are updated into the corpus. The proposed approach is encouraging when compared with other comparative approaches.

Xiaozhi Wang, Ziqi Wang, Xu Han, Wangyi Jiang, Rong Han, Zhiyuan Liu, Juanzi Li, Peng Li, Yankai Lin, Jie Zhou. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020.

Wireless sensor networks may be deployed in many applications to detect and track events of interest. Events can be either point events with an exact location and constant shape, or region events which cover a large area and have dynamic shapes. While both types of events have received attention, no event detection and tracking protocol in existing wireless sensor network research is able to identify and track region events with dynamic shapes(NH-change shapes to identities), which arise when events are created or destroyed through splitting and merging. In this paper, we propose DRAGON, an event detection and tracking protocol which is able to handle all types of events including region events with dynamic identities. DRAGON employs two physics metaphors: event center of mass, to give an approximate location to the event; and node momentum, to guide the detection of event merges and splits. Both detailed theoretical analysis and extensive performance studies of DRAGON's properties demonstrate that DRAGON's execution is distributed among the sensor nodes, has low latency, is energy efficient, is able to run on a wide array of physical deployments, and has performance which scales well with event size, speed, and count.

The task of event detection aims to find the event trigger for a sentence and identify the correct event type. Contextual information is crucial to event detection, which helps a model to identify the triggers better. Existing models utilizing contextual information only take document information as additional features for deep learning model, without considering the specific contribution of document information to trigger classification. In this paper, we propose a Document-Improved Hierarchical Modular Event Detection (DIHMED) model to extract hierarchical contextual information. Specifically, considering the relevance between event types, we build independent modules that combine the document-level information to express this relevance. Given the fact that events from the same document are often related, these modules can make better use of document-level information. We conduct several experiments on a Chinese political event dataset and the results show that our model can outperform the state-of-the-art models.

To compare a positive airway pressure (PAP) device's detection of respiratory events and airway status during device-detected apneas with events scored on simultaneous polysomnography (PSG). Prospective PSGs of patients with sleep apnea using a new-generation PAP device. Four clinical and academic sleep centers. Forty-five patients with obstructive sleep apnea (OSA) and complex sleep apnea (Comp SA) performed a PSG on PAP levels adjusted to induce respiratory events. None. PAP device data identifying the type of respiratory event and whether the airway during a device-detected apnea was open or obstructed were compared to time-synced, manually scored respiratory events on simultaneous PSG recording. Intraclass correlation coefficients between device-detected and PSG scored events were 0.854 for apnea-hypopnea index (AHI), 0.783 for apnea index, 0.252 for hypopnea index, and 0.098 for respiratory event-related arousals index. At a device AHI (AHIFlow) of 10 events/h, area under the receiver operating characteristic curve was 0.98, with sensitivity 0.92 and specificity 0.84. AHIFlow tended to overestimate AHI on PSG at values less than 10 events/h. The device detected that the airway was obstructed in 87.4% of manually scored obstructive apneas. Of the device-detected apneas with clear airway, a minority (15.8%) were manually scored as obstructive apneas. A device-detected apnea-hypopnea index (AHIFlow) < 10 events/h on a positive airway pressure device is strong evidence of good treatment efficacy. Device-detected airway status agrees closely with the presumed airway status during polysomnography scored events, but should not be equated with a specific type of respiratory event.

Event detection (ED) aims to detect events from a given text and categorize them into event types. Most of the current approaches to ED rely heavily on the human annotations of triggers, which are often costly and affect the application of ED in other fields. However, triggers are not necessary for the event detection task. We propose a novel framework called Type Hierarchy Enhanced Event Detection Without Triggers (THEED) to avoid this problem. More specifically, We construct a type hierarchy concept module using the external knowledge graph Probase to enhance the semantic representation of event types. In addition, we divide input instances into word-level and context-level representations, which can make the model use different level features. The experimental result indicates that our proposed approach achieves better improvement. Additionally, it is significantly competitive with mainstream trigger-based models.

Adverse events are unintended patient injuries or complications that arise from healthcare management resulting in death, disability or prolonged hospital stay. Adverse events that require critical care are a considerable financial burden to the healthcare system. Medical record review seems to be a reliable method for detecting adverse events. To synthesize the best available evidence regarding the estimates of the incidence and preventability of adverse events that necessitate intensive care admission; to determine the type and consequences (patient harm, mortality, length of ICU stay and direct medical costs) of these adverse events. MEDLINE (from 1966 to present), EMBASE (from 1974 to present) and CENTRAL (version 1-2010) were searched for studies reporting on unplanned admissions to intensive care units (ICUs). Databases of reports, conference proceedings, grey literature, ongoing research, relevant patient safety organizations and two journals were searched for additional studies. Reference lists of retrieved papers were searched and authors were contacted in an attempt to find any further published or unpublished work. Only quantitative studies that used chart review for the detection of adverse events requiring intensive care admission were considered for eligibility. Studies that were published in the English, Dutch, German, French or Spanish language were included. Two reviewers independently extracted data and assessed the methodological quality of the included studies. 28 studies in the English language and one study in French were included. Of these, two were considered duplicate publications and therefore 27 studies were reviewed. Meta-analysis of the data was not appropriate due to statistical heterogeneity between studies; therefore, results are presented in a descriptive way. Studies were categorized according to the population and the providers of care. 1) The majority of the included studies investigated unplanned intensive care admissions after anesthetic procedures (UIA). 2) Only a few studies examined patients on general wards being at risk for clinical deterioration. The overall incidence of surgical and medical adverse events compared with ICU admissions ranged from 1.1% to 37.2%. 3) The third category of studies examined patients that were readmitted on ICUs. ICU readmission rates varied from 0% to 18.3%. Nine studies explicitly reported on the preventability of adverse outcomes. The preventability rates of the adverse events varied from 17% to 76.5%. Preventable adverse events are further synthesized by type of event and patterns of preventability are being formulated. Consequences of the adverse events included a mean length of ICU stay that ranged from 1.5 days to 10.4 days for the patient's first stay in ICU. Mortality rates varied between 0% and 58%. Adverse events are a persistent and an important reason for admission to the ICU. However, there is relatively weak evidence to estimate an overall incidence and preventability rate of these events. In addition, estimates on preventability are prone to subjective judgments. Variability in methodology and definitions, and poor reporting in studies may be the main reasons for study heterogeneity. Unplanned intensive care admission within 24 hours of a procedure with an anesthetist in attendance (UIA) is a recommended clinical indicator in surgical patients. Several authors recommend early detection of patients with clinical instability on general wards and the implementation of rapid response teams. Step-down or intermediate care units could be a useful strategy for patients that require monitoring to avoid ICU readmissions. There is a need for further studies on the detection of adverse events. The poor quality of current research evidence and the heterogeneity across studies requires that planning of future studies should aim to standardize measures of outcomes to allow for comparisons across studies. This area of research is important in order to identify and explain failure of healthcare systems leading to patient harm, with the ultimate aim to improve the quality and safety of care.

IISD: Integrated Interaction Subgraph Detection for event mining

Social media present a way to discover, report and share different types of events. For this reason, they are considered as a dynamic source of information which enables individuals to stay informed of all real-world events. This specificity encourages researchers to propose methods and approaches to detect events from social media. However many researchers are faced by divers challenges. This is due to the noise of data within the social media. In this paper, we propose a new approach for data retrieval and event detection, which differs from the previous ones by using the mashup concept. The proposed approach is able on the one hand to retrieve data from various social medias which have different structures. On the other hand, our approach aim to detect events with three main dimensions such as the topic, the time and the location.

Detecting events in videos is a complex task, and many different approaches, aimed at a large variety of use-cases, have been proposed in the literature. Most approaches, however, are unimodal and only consider the visual information in the videos. This paper presents and evaluates different approaches based on neural networks where we combine visual features with audio features to detect (spot) and classify events in soccer videos. We employ model fusion to combine different modalities such as video and audio, and test these combinations against different state-of-the-art models on the SoccerNet dataset. The results show that a multimodal approach is beneficial. We also analyze how the tolerance for delays in classification and spotting time, and the tolerance for prediction accuracy, influence the results. Our experiments show that using multiple modalities improves event detection performance for certain types of events.