Hierarchical semi-Markov conditional random fields for deep recursive sequential data

This paper proposes hybrid semi-Markov conditional random fields (SCRFs) for neural sequence labeling in natural language processing. Based on conventional conditional random fields (CRFs), SCRFs have been designed for the tasks of assigning labels to segments by extracting features from and describing transitions between segments instead of words. In this paper, we improve the existing SCRF methods by employing word-level and segment-level information simultaneously. First, word-level labels are utilized to derive the segment scores in SCRFs. Second, a CRF output layer and an SCRF output layer are integrated into a unified neural network and trained jointly. Experimental results on CoNLL 2003 named entity recognition (NER) shared task show that our model achieves state-of-the-art performance when no external knowledge is used.

Automatic semantic annotation of data from databases or the web is an important pre-process for data cleansing and record linkage. It can be used to resolve the problem of imperfect field alignment in a database or identify comparable fields for matching records from multiple sources. The annotation process is not trivial because data values may be noisy, such as abbreviations, variations or misspellings. In particular, overlapping features usually exist in a lexicon-based approach. In this work, we present a probabilistic address parser based on linear-chain conditional random fields (CRFs), which allow more expressive token-level features compared to hidden Markov models (HMMs). In additions, we also proposed two general enhancement techniques to improve the performance. One is taking original semi-structure of the data into account. Another is post-processing of the output sequences of the parser by combining its conditional probability and a score function, which is based on a learned stochastic regular grammar (SRG) that captures segment-level dependencies. Experiments were conducted by comparing the CRF parser to a HMM parser and a semi-Markov CRF parser in two real-world datasets. The CRF parser out-performed the HMM parser and the semi-Markov CRF in both datasets in terms of classification accuracy. Leveraging the structure of the data and combining the linear-chain CRF with the SRG further improved the parser to achieve an accuracy of 97% on a postal dataset and 96% on a company dataset.

Accurately recognizing human activities from sensor data recorded in a smart home setting is a challenging task. Typically, probabilistic models such as the hidden Markov model (HMM) or conditional random fields (CRF) are used to map the observed sensor data onto the hidden activity states. A weakness of these models, however, is that the type of distribution used to model state durations is fixed. Hidden semi-Markov models (HSMM) and semi-Markov conditional random fields (SMCRF) model duration explicitly, allowing state durations to be modelled accurately. In this paper we compare the recognition performance of these models on multiple fully annotated real world datasets consisting of several weeks of data. In our experiments the HSMM consistently outperforms the HMM, showing that accurate duration modelling can result in a significant increase in recognition performance. SMCRFs only slightly outperform CRFs, showing that CRFs are more robust in dealing with violations of the modelling assumptions. The datasets used in our experiments are made available to the community to allow further experimentation.

FLAAP: An Open Human Activity Recognition (HAR) Dataset for Learning and Finding the Associated Activity Patterns

This paper presents a novel framework for the automatic recognition of Activities of Daily Living (ADLs), such as cooking, eating, dishwashing and watching TV, based on depth video processing and Hidden Conditional Random Fields (HCRFs). Depth video is provided by low-cost RGB-D sensors unobtrusively installed in the house. The user’s location, posture, as well as point cloud -based features related to gestures are extracted; a standing/sitting posture detector, as well as novel features expressing head and hand gestures are introduced herein. To model the target activities, we employed discriminative HCRFs and compared them to HMMs. Through experimental evaluation, HCRFs outperformed HMMs in location trajectories-based ADL detection. By fusing trajectories data with posture and the proposed gesture features, ADL detection performance was found to further improve, leading to recognition rates at the level of 90.5 % for five target activities in a naturalistic home environment.KeywordsAdl recognitionUser location trajectoriesPostureGesturesPoint-cloud featuresHidden conditional random fields

For text-query-based keyword spotting from handwritten Chinese documents, the index is usually organized as a candidate lattice to overcome the ambiguity of character segmentation. Each edge in the lattice denotes a candidate character associated with a candidate class. Character similarity (between character and class) scores are calculated on each edge, and the similarity between a query word and handwriting is obtained by combining these edge scores. In this paper, we propose a document indexing method using semi-Markov conditional random fields (semi-CRFs), which provide a principled framework for fusing the information of different contexts. For fast retrieval and to save storage space, the lattice is first purged by a forward-backward pruning approach. On the reduced lattice, we estimate the character similarity scores based on the semi-CRF model. Experimental results on a large handwriting database CASIAOLHWDB justify the effectiveness of the proposed method.

Character-level models of tokens have been shown to be effective at dealing with within-token noise and out-of-vocabulary words. However, they often still rely on correct token boundaries. In this paper, we propose to eliminate the need for tokenizers with an end-to-end character-level semi-Markov conditional random field. It uses neural networks for its character and segment representations. We demonstrate its effectiveness in multilingual settings and when token boundaries are noisy: It matches state-of-the-art part-of-speech taggers for various languages and significantly outperforms them on a noisy English version of a benchmark dataset. Our code and the noisy dataset are publicly available at http://cistern.cis.lmu.de/semiCRF.

In this paper, we proposed a method for linking reviews with database objects problem in data integration, where each object has a set of attributes. To this end, we propose a method based on 2-layer Conditional Random Fields (CRF). First, we show how to apply Semi-Markov CRF to effectively exploit a variety of entity-level features available in integrated data, thereby significantly reducing the dependence on manually labeled training data. Based on the identified entities of the first stage, we link entity reviews with database objects using CRF. Experiments in multiple domains show that our method can substantially superior to traditional tf-idf based methods as well as a recent language model-based method for the review matching problem.

Towards subject independent continuous sign language recognition: A segment and merge approach

In this paper we propose a modified differential evolution (MDE) based feature selection and ensemble learning algorithms for biochemical entity recognizer. Identification and classification of chemical entities are relatively more complex and challenging compared to the other related tasks. As chemical entities we focus on IUPAC and IUPAC related entities. The algorithm performs feature selection within the framework of a robust machine learning algorithm, namely Conditional Random Field. Features are identified and implemented mostly without using any domain specific knowledge and/or resources. In this paper we modify traditional differential evolution to perform two tasks, viz. determining relevant set of features as well as determining proper voting weights for constructing an ensemble. The feature selection technique produces a set of potential solutions on the final population. We develop many models of CRF using these feature combinations. In order to further improve the performance the outputs of these classifiers are combined together using a classifier ensemble technique based on modified DE. Our experiments with the benchmark datasets yield the recall, precision and F-measure values of 82.34%, 88.26% and 85.20%, respectively.KeywordsModified Differential Evolution (MDE)Conditional Random Field (CRF)Feature SelectionEnsembleBiochemical Named Entity

The rapid growing of biological text has promoted the research of the text mining of various non structured documents with the emphasis on the mining of biological knowledge. At the same time, the majority of biological text mining efforts based on the identification of the name of the biological term gene and protein. Therefore, how to recognize the biological terms effectively from the text has become one of the important issues in bioinformatics. Conditional random fields, an important machine learning algorithm, is a model of the probability of a graph model to give an opinion of the label. They traditionally use a set of observations and labels to receive training. Here we use controlled release fertilizer for a class of temporal learning algorithms, in reinforcement learning. Therefore tags are operating, updated environment, the impact of the next observation. Thus, from reinforcement learning, the controlled release fertilizer provides a model of joint action in the decentralized Markov decision process, and defines how agents can communicate with each other, and choose the best way of joint action. We use the hot data corpus for training and testing. The results show that the system can effectively find out the biological terms from the text. We get along with the average accuracy of rate=90.8%, the average recall of rate=90.6%, and the average rate=90.6% F1 six category of biological terms. The results are quite good for the entity recognition system, which is named after many other biological organisms.

The estimation of the Barthel Index scale (BI) is a significant method for measuring the performance of Activities Daily Living (ADL), where the prediction of ADL is crucial for providing rehabilitation care management and recovery for patients after stroke, therefore in this paper, nine various Machine Learning (ML) algorithms were implemented in a medical dataset contains 776 records from 313 patients 208 of them are men: 208 and 150 are women with multiple features collected from them for predicting and classifying the BI status as clinical decision support for determining the ADL of post-stroke patients. Meanwhile, we have applied feature selection using the chi-squared test to reduce the number of features in the dataset. The results showed that the Decision Tree (DT), XGBoost (XGB), and AdaBoost (ADB) classifiers performed the highest performance achieved with 100% correctness in terms of accuracy, sensitivity, specificity, error, and Area Under Curve (AUC) on both the full and reduced features datasets.

Top-down visual saliency facilities object localization by providing a discriminative representation of target objects and a probability map for reducing the search space. In this paper, we propose a novel top-down saliency model that jointly learns a Conditional Random Field (CRF) and a discriminative dictionary. The proposed model is formulated based on a CRF with latent variables. By using sparse codes as latent variables, we train the dictionary modulated by CRF, and meanwhile a CRF with sparse coding. We propose a max-margin approach to train our model via fast inference algorithms. We evaluate our model on the Graz-02 and PASCAL VOC 2007 datasets. Experimental results show that our model performs favorably against the state-of-the-art top-down saliency methods. We also observe that the dictionary update significantly improves the model performance.

We present a discriminative method to classify data that have interdependencies in 2-D lattice. Although both Markov Random Fields (MRFs) and Conditional Random Fields (CRFs) are well-known methods for modeling such dependencies, they are often ineffective and inefficient, respectively. This is because many of the simplifying assumptions that underlie the MRF’s efficiency compromise its accuracy. As CRFs are discriminative, they are typically more accurate than the generative MRFs. This also means their learning process is more expensive. This paper addresses this situation by defining and using “Decoupled Conditional Random Fields (DCRFs)”, a variant of CRFs whose learning process is more efficient as it decouples the tasks of learning potentials. Although our model is only guaranteed to approximate a CRF, our empirical results on synthetic/real datasets show that DCRF is essentially as accurate as other CRF variants, but is many times faster to train.KeywordsSupport Vector MachineMarkov Random FieldConditional Random FieldSequential Minimal OptimizationTumor SegmentationThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Lexis Advance is a legal research service provided by LexisNexis that can respond to natural language queries. It includes a module called Lexis Answers which implements advanced Natural Language Processing (NLP) capabilities to improve understanding of the intent of the user’s queries. Lexis Answers can respond to natural language questions concerning legal question types such as statute of limitations, elements of a claim, definition of legal terms, and others. Herein, we report on the successful use of advanced NLP approaches for detecting not only named entities, but entire legal phrases, a skill previously requiring domain knowledge and human expertise. We have utilized the Conditional Random Fields (CRFs) approach that employs hand-engineered features combined with word2vec embeddings trained on legal corpus. Furthermore, to reduce our dependency on hand-engineered features, we have also implemented deep learning architecture comprising of bidirectional Long Short-Term Memory (BiLSTM) and linear chain CRF. Both approaches were benchmarked against a rule-based approach for different types of legal questions. We find that both CRF and BiLSTM-CRF can identify query intents and legal concepts with comparable precision but much higher recall and F-score than the baseline. The resulting models have been employed in Lexis Answers as critical improvement in our natural language query understanding.