Extract Key Information Research Articles

ISKE: An unsupervised automatic keyphrase extraction approach using the iterated sentences based on graph method

The automatic extraction of key information is necessary for knowledge discovery in this era of rapid knowledge growth. The extraction of key information can also help researchers quickly obtain the information they want instead of reading through all potential documents. Recently, researchers have refocused their attention from words to sentences because utilizing sentences outperforms with respect to illustrating semantics and reduces the calculation complexity. We present a novel and lightweight automatic keyphrase extraction algorithm that does not depend on any external resources, including an external dictionary or corpus. Unlike traditional graph-based algorithms that iterate words to generate keyphrase lists, our proposal uses iterated sentences to rank words and generate keyphrase lists for the semantic information of sentences that are more complete than the word. We initialize the values of words with weighted information and generate a sentence score using these values. Then, we integrate sentences to update their values; hence, the values of the words are updated with the sentence information. We iterate this process until the values of the sentences and words converge. The proposed method is based on a measurement of the relations between sentences and an evaluation of the flow of these relations in an easily understood manner. These relationships are based on the hypothesis that the causality between adjacent sentences is semantically stronger than the causality between words. We not only increase the extraction accuracy, but also reduce the number of iterations of the algorithm. We compare our proposed method with five strong, popular baseline algorithms on four datasets. The results show that our proposed method performs better than the other algorithms on three evaluation metrics.

Construction of the efficient attention prototypical net based on the time–frequency characterization of vibration signals under noisy small sample

The deep learning net has achieved breakthrough results in fault diagnosis. However, most deep learning algorithms rely on massive volumes of data and are difficult to resist the influence of noise. In this study, a new method based on the efficient attention prototypical net (EAPN) is proposed to address the problems. First, frequency slice wavelet transform (FSWT) that can make the signal well described is introduced to convert the signal into images. Subsequently, the efficient lightweight channel attention (ELCA) module and dilated convolution module are created to extract key information across channels under the greater receptive field. Finally, the improved prototypical network is applied to classify based on metric space. Compared with other methods, this paper verified the higher performance of the EAPN under limited samples and strong noise based on the standard bearing fault data set and the actual collected rotor fault data.

On the capability of a class of quantum sensors

Quantum sensors may provide extremely high sensitivity and precision to extract key information in a quantum or classical physical system. A fundamental question is whether a quantum sensor is capable of uniquely inferring unknown parameters in a system for a given structure of the quantum sensor and admissible measurement on the sensor. In this paper, we investigate the capability of a class of quantum sensors which consist of either a single qubit or two qubits. A quantum sensor is coupled to a spin chain system to extract information of unknown parameters in the system. With given initialization and measurement schemes, we employ the similarity transformation approach and the Gröbner basis method to prove that a single-qubit quantum sensor cannot effectively estimate the unknown parameters in the spin chain system while the two-qubit quantum sensor can. The work demonstrates that it is a feasible method to enhance the capability of quantum sensors by increasing the number of qubits in the quantum sensors for some practical applications.

ChEMU 2020: Natural Language Processing Methods Are Effective for Information Extraction From Chemical Patents.

Chemical patents represent a valuable source of information about new chemical compounds, which is critical to the drug discovery process. Automated information extraction over chemical patents is, however, a challenging task due to the large volume of existing patents and the complex linguistic properties of chemical patents. The Cheminformatics Elsevier Melbourne University (ChEMU) evaluation lab 2020, part of the Conference and Labs of the Evaluation Forum 2020 (CLEF2020), was introduced to support the development of advanced text mining techniques for chemical patents. The ChEMU 2020 lab proposed two fundamental information extraction tasks focusing on chemical reaction processes described in chemical patents: (1) chemical named entity recognition, requiring identification of essential chemical entities and their roles in chemical reactions, as well as reaction conditions; and (2) event extraction, which aims at identification of event steps relating the entities involved in chemical reactions. The ChEMU 2020 lab received 37 team registrations and 46 runs. Overall, the performance of submissions for these tasks exceeded our expectations, with the top systems outperforming strong baselines. We further show the methods to be robust to variations in sampling of the test data. We provide a detailed overview of the ChEMU 2020 corpus and its annotation, showing that inter-annotator agreement is very strong. We also present the methods adopted by participants, provide a detailed analysis of their performance, and carefully consider the potential impact of data leakage on interpretation of the results. The ChEMU 2020 Lab has shown the viability of automated methods to support information extraction of key information in chemical patents.

A study on attention-based LSTM for abnormal behavior recognition with variable pooling

Behavior recognition is a well-known computer vision mobile technology. It has been used in many applications such as video surveillance, motion detection on devices, human-computer interaction and sports video, etc. However, most of the existing works ignored the depth and spatio-temporal information so that they resulted in over-fitting and inferior performance. Consequently, a novel framework for behavior recognition is proposed in this paper. In this framework, we propose a target depth estimation algorithm to calculate the 3D spatial position information of the target, and take this information as the input of the behavior recognition model. Simultaneously, in order to obtain more Spatio-temporal information and better handle long-term video, combining with the idea of attention mechanism, we propose a skeleton behavior recognition model which is based on spatio-temporal convolution and attention-based LSTM (ST-CNN & ATT-LSTM). The deep spatial information is merged into each segment, and the model focuses on the key information extraction, which is essential for improving behavior recognition performance. Meanwhile, we use a feature compression method based on variable pooling to solve the problem of inconsistent input sizes caused by multi-person behavior recognition, so that the network can flexibly recognize multi-person skeleton sequences. Finally, the proposed framework is evaluated with real-world surveillance video data, and the results indicate that our framework is superior to existing methods.

Hierarchical graph attention network for relation extraction

Previous research on relation extraction has proved the effectiveness of using dependency trees, which build non-local connections between tokens. However, the existing dependency-based models treat the dependency trees as an inherently flat graph, which causes a loss of dependency information for representing sentences. Besides, they fail to consider the fact that the importance of tokens on the dependency tree varies with different relations to be extracted. In this paper, we propose a novel hierarchical graph attention network HierGAT, which can generate multi-level dependency trees and extract key information from them to improve relation extraction. Specifically, it contains multiple dependency-based graph attention layers, each of which takes a different dependency tree generated by an adaptive subtree pruning strategy as input, and distinguishes the importance of different tokens in a dependency tree. Finally, HierGAT integrates the output token representations of each layer with a multi-head attention mechanism, and learns sentence representation for relation extraction through a pooling layer. The experimental results demonstrate that our method outperforms state-of-the-art baselines on the benchmark datasets.

Undergraduates Graph Interpretation and Scientific Paper Reading Shift from Novice- to Expert-like as a Result of Participation in a Summer Research Experience: A Case Study

Research Experiences for Undergraduate (REU) programs often introduce students to scientific research and STEM career possibilities. However, the program impact on students and their research skill development is not well understood. In a case study with 10 REU students, the authors used eye-tracking and self-report data to determine student strategies for reading scientific papers and interpreting graphs at the beginning and end of the program. The strategies of REU students and science experts were then compared. The REU students changed their strategies and performed more like experts at posttest. These findings indicate that, during the REU, students acquired expert-like strategies necessary to engage with scientific articles and extract key information from graphs. The study demonstrates that eye-tracking can document skill growth in REU students.

A Combinatorial Optimized Knapsack Linear Space for Information Retrieval

Key information extraction can reduce the dimensional effects while evaluating the correct preferences of users during semantic data analysis. Currently, the classifiers are used to maximize the performance of web-page recommendation in terms of precision and satisfaction. The recent method disambiguates contextual sentiment using conceptual prediction with robustness, however the conceptual prediction method is not able to yield the optimal solution. Context-dependent terms are primarily evaluated by constructing linear space of context features, presuming that if the terms come together in certain consumer-related reviews, they are semantically reliant. Moreover, the more frequently they coexist, the greater the semantic dependency is. However, the influence of the terms that coexist with each other can be part of the frequency of the terms of their semantic dependence, as they are non-integrative and their individual meaning cannot be derived. In this work, we consider the strength of a term and the influence of a term as a combinatorial optimization, called Combinatorial Optimized Linear Space Knapsack for Information Retrieval (COLSK-IR). The COLSK-IR is considered as a knapsack problem with the total weight being the “term influence” or “influence of term” and the total value being the “term frequency” or “frequency of term” for semantic data analysis. The method, by which the term influence and the term frequency are considered to identify the optimal solutions, is called combinatorial optimizations. Thus, we choose the knapsack for performing an integer programming problem and perform multiple experiments using the linear space through combinatorial optimization to identify the possible optimum solutions. It is evident from our experimental results that the COLSK-IR provides better results than previous methods to detect strongly dependent snippets with minimum ambiguity that are related to inter-sentential context during semantic data analysis.

SAI-YOLO: A Lightweight Network for Real-Time Detection of Driver Mask-Wearing Specification on Resource-Constrained Devices.

Frequent occurrence and long-term existence of respiratory diseases such as COVID-19 and influenza require bus drivers to wear masks correctly during driving. To quickly detect whether the mask is worn correctly on resource-constrained devices, a lightweight target detection network SAI-YOLO is proposed. Based on YOLOv4-Tiny, the network incorporates the Inception V3 structure, replaces two CSPBlock modules with the RES-SEBlock modules to reduce the number of parameters and computational difficulty, and adds a convolutional block attention module and a squeeze-and-excitation module to extract key feature information. Moreover, a modified ReLU (M-ReLU) activation function is introduced to replace the original Leaky_ReLU function. The experimental results show that SAI-YOLO reduces the number of network parameters and calculation difficulty and improves the detection speed of the network while maintaining certain recognition accuracy. The mean average precision (mAP) for face-mask-wearing detection reaches 86% and the average precision (AP) for mask-wearing normative detection reaches 88%. In the resource-constrained device Raspberry Pi 4B, the average detection time after acceleration is 197 ms, which meets the actual application requirements.

Chinese description generation of dual attention images based on multi-modal fusion

In view of the lack of attention to image features in most of the current research on image Chinese description generation, the low quality of image details, and the low accuracy of generated language description, a multi-modal fusion dual attention image description generation model is proposed. The model uses the pre-trained VGG19 network to extract image features, LSTM extracts key information features of text, and combines image features and text features through attention stitching to obtain the attention weight of each spatial position of the image, and then input it into the LSTM language model Generate descriptive sentences. In the AI Challenger 2017 test set, the BLEU-1 and CIDEr scores of this algorithm can reach 71.2 and 106.1 respectively, which are better than baseline models based on a single attention structure, such as NIC, and perform manual observation and comparison of the effects of several models. The model can generate natural language descriptions that can better represent the details of the picture.

A Novel Energy Management Strategy Design Methodology of a PHEV Based on Data-Driven Approach and Online Signal Analysis

This paper introduces an energy management strategy design method for a plug-in hybrid electric vehicle based on the data-driven approach and online signal analysis. It includes two parts, mode division strategy design, and power distribution strategy design. Using the random forest in data mining technology to analyze optimization results of dynamic programming can quickly extract key information and establish optimal and understandable mode division strategy with high precision and stability directly. Besides, integrating the classic “Engine optimal operating curve control” strategy with wavelet transform and Markov prediction, which not only enhances the adaptability of the strategy to different driving conditions but also improves the fuel economy by reducing the impact of transient power on the engine operation. At the same time, to improve the prediction accuracy of the algorithm without increasing the computational complexity, this paper adds a prediction result correction function to the first-order Markov prediction model to reduce the impact of slow update of the probability matrix on prediction accuracy. The simulation results show the average prediction error of the improved Markov prediction model is reduced by 5.3% and the new energy management strategy designed reduces fuel consumption by 8.28% at the cost of a small increase in electricity consumption.

Integrally Cooperative Spatio-Temporal Feature Representation of Motion Joints for Action Recognition

In contemporary research on human action recognition, most methods separately consider the movement features of each joint. However, they ignore that human action is a result of integrally cooperative movement of each joint. Regarding the problem, this paper proposes an action feature representation, called Motion Collaborative Spatio-Temporal Vector (MCSTV) and Motion Spatio-Temporal Map (MSTM). MCSTV comprehensively considers the integral and cooperative between the motion joints. MCSTV weighted accumulates limbs’ motion vector to form a new vector to account for the movement features of human action. To describe the action more comprehensively and accurately, we extract key motion energy by key information extraction based on inter-frame energy fluctuation, project the energy to three orthogonal axes and stitch them in temporal series to construct the MSTM. To combine the advantages of MSTM and MCSTV, we propose Multi-Target Subspace Learning (MTSL). MTSL projects MSTM and MCSTV into a common subspace and makes them complement each other. The results on MSR-Action3D and UTD-MHAD show that our method has higher recognition accuracy than most existing human action recognition algorithms.

DB-Net: Dual Attention Network with Bilinear Pooling for Fire-Smoke Image Classification

In recent years, with the help of deep learning, image classification has been significantly improved. More and more researchers are seeking fire alarm solutions based on these artificial intelligence algorithms. However, this raises new challenges in an effective method of classification for fire-smoke images. In this paper, we propose a Dual Attention Network with Bilinear Pooling of three steps to accomplish the task of fire-smoke classification. We first extract features of fire pictures or smoke ones by using ResNet-50 as a basic net. Then we attach two kinds of attention modules, channel attention and spatial attention, which are called dual attention, to extract key information ‘what’ and ‘where’ from pictures. Finally, we merge them by using the bilinear pooling module which has been shown to be effective at improving our classification rate. Results show that our most accurate model can reach 90.11% per-image accuracy, which is improved by 4.81% compared to the traditional ResNet-50.

Characterization of the Topological Features of Catalytic Sites in Protein Coevolution Networks**Supported by the National Natural Science Foundation of China (Grant No. 11305139).

The knowledge of sequence and structural properties of residues in the catalytic sites of enzymes is important for understanding the physiochemical basis of enzymatic catalysis. We reveal new features of the catalytic sites by analyzing the coevolutionary behavior of amino acid sequences. By performing direct coupling analysis of the sequences of homologous proteins, we construct the coevolution networks at the residue level. Based on the analysis of the topological features of the coevolution networks for a dataset including 20 enzymes, we show that there is significant correlation between the catalytic sites and topological features of protein coevolution networks. Residues at the catalytic center often correspond to the nodes with high values of centralities in the networks as characterized by the degree, betweenness, closeness, and Laplacian centrality. The results of this work provide a possible way to extract key coevolutionary information from the sequences of enzymes, which is useful in the prediction of catalytic sites of enzymes.

4541 Harnessing the Power of the Electronic Medical Record in Interstitial Lung Disease

OBJECTIVES/GOALS: Harnessing the EHR to support clinical research is critical for the study of rare diseases such as interstitial lung disease (ILD). However no studies have compared EHR and research-quality data in the ILD population. Our objectives were to (1) identify ILD patients and extract clinical data from an EHR system and (2) assess the performance of ILD data capture. METHODS/STUDY POPULATION: Case validated algorithms were implemented to identify patients from the University of California San Francisco EHR and extract key ILD clinical information including, demographic variables, process measures and patient outcomes. Key clinical information were defined based on consensus statements and ILD clinical trials. A subset of ILD patients, had variables recorded in both the EHR and a separate ILD longitudinal research database. The completeness of EHR data capture and level of agreement were compared between three data collection methods: (1) data manually and systematically collected for an ILD research database (gold standard), (2) data automatically extracted from structured fields in the EHR, and (3) data extracted from unstructured data sources. RESULTS/ANTICIPATED RESULTS: We identified 5857 ILD patients in the EHR, of which 2100 patients had data available in the both the EHR and research database. Baseline demographic variables, co-morbidities, use of diagnostic testing, pharmacotherapy were accurately extracted from structured fields. Outcome measures, including lung physiology, radiographic patterns, pathology results, and health related quality of life (HRQoL) were unevenly extracted from structured fields alone. With the exception of HRQoL, these measures were accurately captured in unstructured EHR sources. Notably, certain metrics were better defined in the EHR, including health care resource utilization metrics, acute exacerbations, medication side effects, supplemental oxygen use and specialty care referrals (rheumatology, lung transplant, palliative care, etc). DISCUSSION/SIGNIFICANCE OF IMPACT: A large real-world ILD cohort can be algorithmically extracted from the EHR along with key clinical variables with accuracy comparable to protocol-driven research databases. Rigorous assessment of the types of disease-specific variables that are present in EHR-derived data will inform future interventions to improve the fidelity, accessibility and use of the EHR in clinical research.

Analysis of COVID-19 spread characteristics and infection numbers based on large-scale structured case data

In early 2020, the novel coronavirus, referred to as COVID-19 burst out. The Chinese people took the most comprehensive and rigorous control measures to fight against the COVID-19. Local health control departments reported infection data in a timely manner, which helped the public understand the development of the epidemic and take protective measures in advance. However, currently, no literature has analyzed the transmission characteristics of COVID-19 based on the structured data of large-scale patient cases and artificial intelligence. The detailed case data of patients in various regions are primarily recorded in text form, and the formats of report data in different provinces and cities differ, which makes it difficult to handle such data. To analysis around a large anonymous patient case data, we propose a method based on natural language processing technology to structure the case data. The proposed method can extract key information in the cases accurately and effectively with the help a pretrained model and a small number of labeled samples. By mining the patient's structured case data, we analyze the gender and age distribution, the main causes of infection, the characteristics of the incubation period, and epidemic trends in detail. Using big data on travel, a method was developed to estimate the number of infected individuals in Wuhan prior the restrictions were put into effect. This method helps people understand the real epidemic situation and take execute early protective measures. It is also helps government departments make evidence-based decisions, dispatch medical staff, and allocate medical resources as quickly as possible.

Chinese long text summarization using improved sequence-to-sequence lstm

Text summarization is an important issue in natural language processing. The existing method has the problem of low accuracy when performing long text summarization. In this paper, We use the LSTM to construct the sequence-to-sequence model, and combine the attention mechanism to process automatic Chinese long text summarization.The experimental results indicate that our method can accurately extract key information from long text, generate high-quality summary.

Fault classification in three-phase motors based on vibration signal analysis and artificial neural networks

Competition in the industrial environment is increasingly intense, so it is of utmost importance that organizations keep their assets in operation as much as possible (in order to produce more). In this context, there is a need for predictive maintenance, a technique that detects the health of assets in real time, allowing failures to be diagnosed before they can interrupt the operation of the assets, avoiding high financial losses. This study uses a sixteen-motor experimental setup with four different known operating conditions. The vibration signal of these motors, through signal analysis, both in time and frequency domains, is performed to evaluate the types and severities of the defects. An artificial neural network (ANN) is used to classify these defects. Considering the vibration analysis, mechanical faults can be identified quickly and conveniently. For the development of the ANN, it was necessary to perform a preprocessing of the vibration signal (response in time) due to the data size, which overwhelms the network. Thus, statistical data were used to extract key information from the vibration signal. Finally, the neural network created based on this study’s methodology presents extremely reliable results, allowing a quick and robust diagnosis of the motor operating condition.

Matching patients to clinical trials using semantically enriched document representation.

Recruiting eligible patients for clinical trials is crucial for reliably answering specific questions about medical interventions and evaluation. However, clinical trial recruitment is a bottleneck in clinical research and drug development. Our goal is to provide an approach towards automating this manual and time-consuming patient recruitment task using natural language processing and machine learning techniques. Specifically, our approach extracts key information from series of narrative clinical documents in patient's records and collates helpful evidence to make decisions on eligibility of patients according to certain inclusion and exclusion criteria. Challenges in applying narrative clinical documents such as differences in reporting styles and sub-languages are addressed by enriching them with knowledge from domain ontologies in the form of semantic vector representations. We show that a machine learning model based on Multi-Layer Perceptron (MLP) is more effective for the task than five other neural networks and four conventional machine learning models. Our approach achieves overall micro-F1-Score of 84% for 13 different eligibility criteria. Our experiments also indicate that semantically enriched documents are more effective than using original documents for cohort selection. Our system provides an end-to-end machine learning-based solution that achieves comparable results with the state-of-the-art which relies on hand-crafted rules or data-centric engineered features.

Dynamic event type recognition and tagging for data-driven insights in law-enforcement

In law enforcement, investigators are typically tasked with analyzing large collections of evidences in order to identify and extract key information to support investigation cases. In this context, events are key elements that help understanding and reconstructing what happened from the collection of evidence items. With the ever increasing amount of data (e.g., e-mails and content from social media) gathered today as part of investigation tasks (in most part done manually), managing such amount of data can be challenging and prone to missing important details that could be of high relevance to a case. In this paper, we aim to facilitate the work of investigators through a framework for deriving insights from data. We focus on the auto-recognition and dynamic tagging of event types (e.g., phone calls) from (textual) evidence items, and propose a framework to facilitate these tasks and provide support for insights and discovery. The experimental results obtained by applying our approach to a real, legal dataset demonstrate the feasibility of our proposal by achieving good performance in the task of automatically recognizing and tagging event types of interest.