Conditional Random Field Classifier Research Articles

Voting Strategies for Arabic Named Entity Recognition using Annotation Schemes

Named Entity Recognition (NER) seeks to identify and classify NEs into predefined categories and is an important subtask in information extraction. Many annotation schemes have been proposed to assign suitable labels for multiword NEs within a given text. This study proposes a method to combine the results of different annotation schemes (IOB, IOE, IOBE, IOBS, IOES, and IOBES) for Arabic NER (ANER). Three voting strategies are explored, namely, majority voting, weighted voting, and weighted voting-based Particle Swarm Optimization (PSO), applied to Conditional Random Fields (CRF) classifiers, each corresponding to a certain annotation scheme. The experimental results showed that majority voting can be considered an effective combination strategy to enhance the performance of ANER systems.

Multi-head CRF classifier for biomedical multi-class named entity recognition on Spanish clinical notes.

The identification of medical concepts from clinical narratives has a large interest in the biomedical scientific community due to its importance in treatment improvements or drug development research. Biomedical named entity recognition (NER) in clinical texts is crucial for automated information extraction, facilitating patient record analysis, drug development, and medical research. Traditional approaches often focus on single-class NER tasks, yet recent advancements emphasize the necessity of addressing multi-class scenarios, particularly in complex biomedical domains. This paper proposes a strategy to integrate a multi-head conditional random field (CRF) classifier for multi-class NER in Spanish clinical documents. Our methodology overcomes overlapping entity instances of different types, a common challenge in traditional NER methodologies, by using a multi-head CRF model. This architecture enhances computational efficiency and ensures scalability for multi-class NER tasks, maintaining high performance. By combining four diverse datasets, SympTEMIST, MedProcNER, DisTEMIST, and PharmaCoNER, we expand the scope of NER to encompass five classes: symptoms, procedures, diseases, chemicals, and proteins. To the best of our knowledge, these datasets combined create the largest Spanish multi-class dataset focusing on biomedical entity recognition and linking for clinical notes, which is important to train a biomedical model in Spanish. We also provide entity linking to the multi-lingual Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT) vocabulary, with the eventual goal of performing biomedical relation extraction. Through experimentation and evaluation of Spanish clinical documents, our strategy provides competitive results against single-class NER models. For NER, our system achieves a combined micro-averaged F1-score of 78.73, with clinical mentions normalized to SNOMED CT with an end-to-end F1-score of 54.51. The code to run our system is publicly available at https://github.com/ieeta-pt/Multi-Head-CRF. Database URL: https://github.com/ieeta-pt/Multi-Head-CRF.

A Study on the Performance of Recurrent Neural Network based Models in Maithili Part of Speech Tagging

This article presents our effort in developing a Maithili Part of Speech (POS) tagger. Substantial effort has been devoted to developing POS taggers in several Indian languages, including Hindi, Bengali, Tamil, Telugu, Kannada, Punjabi, and Marathi; but Maithili did not achieve much attention from the research community. Maithili is one of the official languages of India, with around 50 million native speakers. So, we worked on developing a POS tagger in Maithili. For the development, we use a manually annotated in-house Maithili corpus containing 56,126 tokens. The tagset contains 27 tags. We train a conditional random fields (CRF) classifier to prepare a baseline system that achieves an accuracy of 82.67%. Then, we employ several recurrent neural networks (RNN)-based models, including Long-short Term Memory (LSTM), Gated Recurrent Unit (GRU), LSTM with a CRF layer (LSTM-CRF), and GRU with a CRF layer (GRU-CRF) and perform a comparative study. We also study the effect of both word embedding and character embedding in the task. The highest accuracy of the system is 91.53%.

Tagging terms in text

Abstract As with many tasks in natural language processing, automatic term extraction (ATE) is increasingly approached as a machine learning problem. So far, most machine learning approaches to ATE broadly follow the traditional hybrid methodology, by first extracting a list of unique candidate terms, and classifying these candidates based on the predicted probability that they are valid terms. However, with the rise of neural networks and word embeddings, the next development in ATE might be towards sequential approaches, i.e., classifying each occurrence of each token within its original context. To test the validity of such approaches for ATE, two sequential methodologies were developed, evaluated, and compared: one feature-based conditional random fields classifier and one embedding-based recurrent neural network. An additional comparison was added with a machine learning interpretation of the traditional approach. All systems were trained and evaluated on identical data in multiple languages and domains to identify their respective strengths and weaknesses. The sequential methodologies were proven to be valid approaches to ATE, and the neural network even outperformed the more traditional approach. Interestingly, a combination of multiple approaches can outperform all of them separately, showing new ways to push the state-of-the-art in ATE.

Context-aware sequence labeling for condition information extraction from historical bridge inspection reports

Effective upkeep of aging infrastructure systems with limited funding and resources calls for efficient bridge management systems. Although data-driven models have been extensively studied in the last decade for extracting knowledge from past experience to guide future maintenance decision making, their performance and usefulness have been limited by the level of detail and accuracy of currently available bridge condition databases. This paper leverages an untapped resource for bridge condition data and proposes a new method to extract condition information from it at a high level of detail. To that end, a natural language processing approach was developed to formalize structural condition knowledge by formulating a sequence labeling task and modeling inspection narratives as a combination of words representing defects, their severity and location, while accounting for the context of each word. The proposed framework employs a deep-learning-based approach and incorporates context-aware components including a bi-directional Long Short Term Memory (LSTM) neural network architecture and a Conditional Random Field (CRF) classifier to account for the context of words when assigning labels. A dependency-based word embedding model was also used to represent the raw text while incorporating both semantic and contextual information. The sequence labeling model was trained using bridge inspection reports collected from the Virginia Department of Transportation bridge inspection database and achieved an F1 score of 94.12% during testing. The proposed model also demonstrated improvements compared with baseline sequence labeling models, and was further used to demonstrate the capability of detecting condition changes with respect to previous inspection records. Results of this study show that the proposed method can be used to extract and create a condition information database that can further assist in developing data-driven bridge management and condition forecasting models, as well as automated bridge inspection systems.

WHU-Hi: UAV-borne hyperspectral with high spatial resolution (H2) benchmark datasets and classifier for precise crop identification based on deep convolutional neural network with CRF

Unmanned aerial vehicle (UAV)-borne hyperspectral systems can acquire hyperspectral imagery with a high spatial resolution (which we refer to here as H2 imagery). As a result of the low operating cost, high flexibility, and the ability to achieve real-time data acquisition, UAV-borne hyperspectral systems have become an important data source for remote sensing based agricultural monitoring. However, precise crop classification based on UAV-borne H2 imagery is a challenging task when faced with a number of different crop classes. The traditional hyperspectral classification methods, such as the spectral-based and object-oriented classification methods, have difficulty in classifying H2 imagery, faced with the problems of salt-and-pepper (SP) noise and scale selection. In this article, the deep convolutional neural network with a conditional random field classifier (CNNCRF) framework is proposed for precise crop classification with UAV-borne H2 imagery. In the proposed method, a deep convolutional neural network (CNN) is designed to extract and fuse in-depth spectral and local spatial features, and the conditional random field (CRF) model further incorporates the spatial-contextual information to improve the problem of holes and isolated regions in the classification map. Meanwhile, virtual sample augmentation based on the hyperspectral imaging mechanism is used to lessen the issue of the limited labeled samples. To validate the results, a new dataset—the Wuhan UAV-borne hyperspectral image (WHU-Hi) dataset—has been built for precise crop classification. The experimental results obtained using the WHU-Hi dataset confirm the accuracy and visualization performance of the proposed CNNCRF classification method, which outperforms the previous methods. In addition, the WHU-Hi dataset could serve as a benchmark dataset for hyperspectral image classification studies.

Towards the first Maithili part of speech tagger: Resource creation and system development

Part of speech (POS) tagging for the Indian language Maithili is not an explored territory. There have been substantial efforts at developing POS taggers in several Indian languages including Hindi, Bengali, Tamil, Telugu, Kannada, Punjabi and Marathi; but we did not find any openly available POS tagger and tagged corpus in Maithili. However, Maithili is one of the official languages of India with around 50 million native speakers. Development of Maithili natural language processing (NLP) tools and resources is extremely important as the language is currently being used in education and official contexts in certain states in India. In this paper, we present our effort on the development of a Maithili POS tagger. As we did not find any open training data, we started the development by annotation of a POS tagged corpus. We defined a POS tagset and manually annotated a Maithili corpus containing 52,190 words. We used the corpus to train a conditional random fields (CRF) classifier. We ran experiments using various feature sets and achieved an accuracy of 82.67%. Then we collected large raw corpora containing Wikipedia dump and other Maithili web resources to train neural word embedding. The word2vec CBOW model was trained and the generated word vectors were utilized during CRF training. With this inclusion, the accuracy of the system increased to 85.88%.

Sequence labeling to detect stuttering events in read speech

Stuttering is a speech disorder that, if treated during childhood, may be prevented from persisting into adolescence. A clinician must first determine the severity of stuttering, assessing a child during a conversational or reading task, recording each instance of disfluency, either in real time, or after transcribing the recorded session and analysing the transcript. The current study evaluates the ability of two machine learning approaches, namely conditional random fields (CRF) and bi-directional long-short-term memory (BLSTM), to detect stuttering events in transcriptions of stuttering speech. The two approaches are compared for their performance both on ideal hand-transcribed data and also on the output of automatic speech recognition (ASR). We also study the effect of data augmentation to improve performance. A corpus of 35 speakers’ read speech (13K words) was supplemented with a corpus of 63 speakers’ spontaneous speech (11K words) and an artificially-generated corpus (50K words). Experimental results show that, without feature engineering, BLSTM classifiers outperform CRF classifiers by 33.6%. However, adding features to support the CRF classifier yields performance improvements of 45% and 18% over the CRF baseline and BLSTM results, respectively. Moreover, adding more data to train the CRF and BLSTM classifiers consistently improves the results.

Aspect Term Extraction for Aspect Based Opinion Mining

Opinion Mining (OM) is also called as Sentiment Analysis (SA). Aspect Based Opinion Mining (ABOM) is also called as Aspect Based Sentiment Analysis (ABSA). In this paper, three new features are proposed to extract the aspect term for Aspect Based Sentiment Analysis (ABSA). The influence of the proposed features is evaluated on five classifiers namely Decision Tree (DT), Naive Bayes (NB), K-Nearest Neighbor (KNN), Support Vector Machine (SVM) and Conditional Random Fields (CRF). The proposed features are evaluated on the Two datasets on Restaurant and Laptop domains available in International Workshop on Semantic Evaluation 2014 i.e. SemEval 2014. The influence of proposed features is evaluated using Precision, Recall and F1 measures. The proposed features are highly influencing for aspect term extraction on classifiers. The performance of SVM and CRF classifiers with proposed features is more influencing for aspect term extraction compared with NB, DT and KNN classifiers.

Prediction of human driver intentions at a narrow passage in inner city traffic / Intentionsprädiktion menschlicher Fahrer an einer Engstelle im innerstädtischen Straßenverkehr

Abstract Autonomous vehicles have to be able to predict whether a human driver will wait at an unregulated inner city narrow passage or not to adapt its behaviour accordingly. To this end, a driving simulator study was conducted in which participiants were subjected to different cooperation behaviours during their approach to a narrow passage. They were asked to rate their intention afterwards. From the recorded trajectories, features which are specific to the scenario are derived. Therewith, Random Forest and Conditional Random Field classifiers for both intention and behaviour prediction are trained. The results show that robust prediction of driver intention and behaviour is possible.

A COMPARATIVE STUDY OF WORD REPRESENTATION METHODS WITH CONDITIONAL RANDOM FIELDS AND MAXIMUM ENTROPY MARKOV FOR BIO-NAMED ENTITY RECOGNITION

Bio-Named Entity Recognition (Bio-NER) is the process of identifying and semantically classifying biomedical technical terms and named entities in Biomedicine literature. Therefore, it is a major task in biomedical knowledge acquisition. Meanwhile, Natural Language Processing (NLP) plays an important role in Bio-NER in the biomedical domain. The first and most essential biomedical literature mining task incorporates biomedical entity recognition such as protein, gene, and chemicals. The most recent Bio-NER methods rely on predefined traditional features, which attempt to capture the specific surface properties of entity types. However, these empirically predefined feature sets differ between entity types and are manually constructed and complicated, which means developing them is costly. In this paper, we systematically present a comparative evaluation study of three methods, which are: the traditional feature representation method, the continuous bag-of-words (CBOW) model, and a new prototypical representation method with two popular sequence-labeling approaches (Conditional Random Fields (CRFs) and Maximum Entropy Markov Models (MEMM)). We evaluated these models with two major Bio-NER tasks, which involve the JNLPBA and GENETAG corpora. This paper examined the prototypical word representation method and found that Word2Vec can be successfully used for Bio-NER. Our results show that the new prototypical representation method improved the performance of the two machine learning models with different datasets. Also, the new prototypical representation method performed better than the traditional feature representation method and CBOW model for both datasets. Finally, our experiment proved that the CRF classifier with the new prototypical representation method achieved the best results when 90% data was used as training data, yielding overall F-measure values of 0.79% and 0.85% for the JNLPBA corpus and GENETAG corpus, respectively. In comparison, the results achieved using the ME classifier yielded overall F-measure values of 0.76% and 0.78% for the JNLPBA corpus and GENETAG corpus, respectively.

A Low-Cost Named Entity Recognition Research Based on Active Learning

Named entity recognition (NER) is an indispensable and very important part of many natural language processing technologies, such as information extraction, information retrieval, and intelligent Q & A. This paper describes the development of the AL-CRF model, which is a NER approach based on active learning (AL). The algorithmic sequence of the processes performed by the AL-CRF model is the following: first, the samples are clustered using the k-means approach. Then, stratified sampling is performed on the produced clusters in order to obtain initial samples, which are used to train the basic conditional random field (CRF) classifier. The next step includes the initiation of the selection process which uses the criterion of entropy. More specifically, samples having the highest entropy values are added to the training set. Afterwards, the learning process is repeated, and the CRF classifier is retrained based on the obtained training set. The learning and the selection process of the AL is running iteratively until the harmonic mean F stabilizes and the final NER model is obtained. Several NER experiments are performed on legislative and medical cases in order to validate the AL-CRF performance. The testing data include Chinese judicial documents and Chinese electronic medical records (EMRs). Testing indicates that our proposed algorithm has better recognition accuracy and recall rate compared to the conventional CRF model. Moreover, the main advantage of our approach is that it requires fewer manually labelled training samples, and at the same time, it is more effective. This can result in a more cost effective and more reliable process.

Comprehensive Word-Level Classification of Screening Mammography Reports Using a Neural Network Sequence Labeling Approach.

Radiology reports contain a large amount of potentially valuable unstructured data. Recently, neural networks have been employed to perform classification of radiology reports over a few classes at the document level. The success of neural networks in sequence-labeling problems such as named entity recognition and part of speech tagging suggests that they could be used to classify radiology report text with greater granularity. We employed a neural network architecture to comprehensively classify mammography report text at the word level using a sequence labeling approach. Two radiologists devised a comprehensive classification system for screening mammography reports. Each word in each report was manually categorized by a radiologist into one of 33 categories according to the classification system. Tagged words referencing the same finding were grouped into unique sets. We pre-labeled reports with a rule-based algorithm and then manually edited these annotations for 6705 screening mammography reports (25.1%, 66.8%, and 8.1% BI-RADS 0, 1, and 2, respectively). A combined convolutional and recurrent neural network model was used to label words in each sentence of the individual reports. A siamese recurrent neural network was then used to group findings into sets. Performance of the neural network-based method was compared to a rule-based algorithm and a conditional random field (CRF) model. Global accuracy (percentage of documents where all word tags were predicted correctly) and keyword accuracy (percentage of all words that were labeled correctly, excluding words tagged as unimportant) were calculated on an unseen 519 report test set. Two-tailed t tests were used to assess differences between algorithm performance, and p< 0.05 was used to determine statistical significance. The neural network-based approach showed significantly higher global accuracy compared to both the rule-based algorithm (88.3 vs 57.0%, p< 0.001) and the CRF model (88.3% vs. 75.8%, p< 0.001). The neural network also showed significantly higher keyword level accuracy compared to the rule-based algorithm (95.5% vs. 80.9% p< 0.001) and CRF model (95.5% vs. 76.9%, p< 0.001). We demonstrate the potential of neural networks to accurately perform word-level multilabel classification of free text radiology reports across 33 classes, thus showing the utility of a sequence labeling approach to NLP of radiology reports. We found that a neural network classifier outperforms a rule-based algorithm and a CRF classifier for comprehensive multilabel classification of free text screening mammography reports at the word level. By approaching radiology report classification as a sequence-labeling problem, we demonstrate the ability of neural networks to extract data from free text radiology reports at a level of granularity not previously reported.

Generative vs. Discriminative Recognition Models for Off-Line Arabic Handwriting.

The majority of handwritten word recognition strategies are constructed on learning-based generative frameworks from letter or word training samples. Theoretically, constructing recognition models through discriminative learning should be the more effective alternative. The primary goal of this research is to compare the performances of discriminative and generative recognition strategies, which are described by generatively-trained hidden Markov modeling (HMM), discriminatively-trained conditional random fields (CRF) and discriminatively-trained hidden-state CRF (HCRF). With learning samples obtained from two dissimilar databases, we initially trained and applied an HMM classification scheme. To enable HMM classifiers to effectively reject incorrect and out-of-vocabulary segmentation, we enhance the models with adaptive threshold schemes. Aside from proposing such schemes for HMM classifiers, this research introduces CRF and HCRF classifiers in the recognition of offline Arabic handwritten words. Furthermore, the efficiencies of all three strategies are fully assessed using two dissimilar databases. Recognition outcomes for both words and letters are presented, with the pros and cons of each strategy emphasized.

Deep Learning-Based Language Identification in English-Hindi-Bengali Code-Mixed Social Media Corpora

Abstract This article addresses language identification at the word level in Indian social media corpora taken from Facebook, Twitter and WhatsApp posts that exhibit code-mixing between English-Hindi, English-Bengali, as well as a blend of both language pairs. Code-mixing is a fusion of multiple languages previously mainly associated with spoken language, but which social media users also deploy when communicating in ways that tend to be rather casual. The coarse nature of code-mixed social media text makes language identification challenging. Here, the performance of deep learning on this task is compared to feature-based learning, with two Recursive Neural Network techniques, Long Short Term Memory (LSTM) and bidirectional LSTM, being contrasted to a Conditional Random Fields (CRF) classifier. The results show the deep learners outscoring the CRF, with the bidirectional LSTM demonstrating the best language identification performance.

Using long short-term memory deep neural networks for aspect-based sentiment analysis of Arabic reviews

This paper proposes a state-of-the-art research for aspect-based sentiment analysis of Arabic Hotels’ reviews using two implementations of long short-term memory (LSTM) neural networks. The first one is (a) a character-level bidirectional LSTM along with conditional random field classifier (Bi-LSTM-CRF) for aspect opinion target expressions (OTEs) extraction, and the second one is (b) an aspect-based LSTM for aspect sentiment polarity classification in which the aspect-OTEs are considered as attention expressions to support the sentiment polarity identification. Proposed approaches are evaluated using a reference dataset of Arabic Hotels’ reviews. Results show that our approaches outperform baseline research on both tasks with an enhancement of 39% for the task of aspect-OTEs extraction and 6% for the aspect sentiment polarity classification task.

Machine Reading of Biomedical Data Dictionaries

This article describes an approach for the automated reading of biomedical data dictionaries. Automated reading is the process of extracting element details for each of the data elements from a data dictionary in a document format (such as PDF) to a completely structured representation. A structured representation is essential if the data dictionary metadata are to be used in applications such as data integration and also in evaluating the quality of the associated data. We present an approach and implemented solution for the problem, considering different formats of data dictionaries. We have a particular focus on the most challenging format with a machine-learning classification solution to the problem using conditional random field classifiers. We present an evaluation using several actual data dictionaries, demonstrating the effectiveness of our approach.

SEMANTIC LABELLING OF ULTRA DENSE MLS POINT CLOUDS IN URBAN ROAD CORRIDORS BASED ON FUSING CRF WITH SHAPE PRIORS

Abstract. In this paper, a labelling method for the semantic analysis of ultra-high point density MLS data (up to 4000 points/m2) in urban road corridors is developed based on combining a conditional random field (CRF) for the context-based classification of 3D point clouds with shape priors. The CRF uses a Random Forest (RF) for generating the unary potentials of nodes and a variant of the contrastsensitive Potts model for the pair-wise potentials of node edges. The foundations of the classification are various geometric features derived by means of co-variance matrices and local accumulation map of spatial coordinates based on local neighbourhoods. Meanwhile, in order to cope with the ultra-high point density, a plane-based region growing method combined with a rule-based classifier is applied to first fix semantic labels for man-made objects. Once such kind of points that usually account for majority of entire data amount are pre-labeled; the CRF classifier can be solved by optimizing the discriminative probability for nodes within a subgraph structure excluded from pre-labeled nodes. The process can be viewed as an evidence fusion step inferring a degree of belief for point labelling from different sources. The MLS data used for this study were acquired by vehicle-borne Z+F phase-based laser scanner measurement, which permits the generation of a point cloud with an ultra-high sampling rate and accuracy. The test sites are parts of Munich City which is assumed to consist of seven object classes including impervious surfaces, tree, building roof/facade, low vegetation, vehicle and pole. The competitive classification performance can be explained by the diverse factors: e.g. the above ground height highlights the vertical dimension of houses, trees even cars, but also attributed to decision-level fusion of graph-based contextual classification approach with shape priors. The use of context-based classification methods mainly contributed to smoothing of labelling by removing outliers and the improvement in underrepresented object classes. In addition, the routine operation of a context-based classification for such high density MLS data becomes much more efficient being comparable to non-contextual classification schemes.

A Novel Word Clustering and Cluster Merging Technique for Named Entity Recognition

Abstract In this paper, we present a novel word clustering technique to capture contextual similarity among the words. Related word clustering techniques in the literature rely on the statistics of the words collected from a fixed and small word window. For example, the Brown clustering algorithm is based on bigram statistics of the words. However, in the sequential labeling tasks such as named entity recognition (NER), longer context words also carry valuable information. To capture this longer context information, we propose a new word clustering algorithm, which uses parse information of the sentences and a nonfixed word window. This proposed clustering algorithm, named as variable window clustering, performs better than Brown clustering in our experiments. Additionally, to use two different clustering techniques simultaneously in a classifier, we propose a cluster merging technique that performs an output level merging of two sets of clusters. To test the effectiveness of the approaches, we use two different NER data sets, namely, Hindi and BioCreative II Gene Mention Recognition. A baseline NER system is developed using conditional random fields classifier, and then the clusters using individual techniques as well as the merged technique are incorporated to improve the classifier. Experimental results demonstrate that the cluster merging technique is quite promising.

WEAKLY SUPERVISED SEGMENTATION-AIDED CLASSIFICATION OF URBAN SCENES FROM 3D LIDAR POINT CLOUDS

Abstract. We consider the problem of the semantic classification of 3D LiDAR point clouds obtained from urban scenes when the training set is limited. We propose a non-parametric segmentation model for urban scenes composed of anthropic objects of simple shapes, partionning the scene into geometrically-homogeneous segments which size is determined by the local complexity. This segmentation can be integrated into a conditional random field classifier (CRF) in order to capture the high-level structure of the scene. For each cluster, this allows us to aggregate the noisy predictions of a weakly-supervised classifier to produce a higher confidence data term. We demonstrate the improvement provided by our method over two publicly-available large-scale data sets.