Considering the non-linear properties of the human visual system, many non-linear operators and models have been developed, particularly the logarithmic image processing (LIP) model proposed by Jourlin and Pinoli, which has been proved to be physically justified in several laws of the human visual system and has been successfully applied in image processing areas. Recently, several modifications based on this logarithmic mathematical framework have been presented, such as parameterized logarithmic image processing (PLIP), pseudo-logarithmic image processing, homomorphic logarithmic image processing. In this paper, a new single parameter logarithmic model for image processing with an adaptive parameter-based Sobel edge detection algorithm is presented. On the basis of analyzing the distributive law, the subtractive law, and the isomorphic property of the PLIP model, the five parameters in PLIP are replaced by a single parameter to ensure the completeness of the model and physical constancy with the nature of an image, and then an adaptive parameter-based Sobel edge detection algorithm is proposed. By using an image noise estimation method to evaluate the noise level of image, the adaptive parameter in the single parameter LIP model is calculated based on the noise level and grayscale value of a corresponding image area, followed by the single-parameter LIP-based Sobel operation to overcome the noise-sensitive problem of classical LIP-based Sobel edge detection methods, especially in the dark area of an image, while retaining edge sensitivity. Compared with the classical LIP and PLIP model, the given single parameter LIP achieves satisfactory results in noise suppression and edge accuracy.

In this paper, we propose a distributed speaker recognition method using a nonparametric speaker model and Earth Mover's Distance (EMD). In distributed speaker recognition, the quantized feature vectors are sent to a server. The Gaussian mixture model (GMM), the traditional method used for speaker recognition, is trained using the maximum likelihood approach. However, it is difficult to fit continuous density functions to quantized data. To overcome this problem, the proposed method represents each speaker model with a speaker-dependent VQ code histogram designed by registered feature vectors and directly calculates the distance between the histograms of speaker models and testing quantized feature vectors. To measure the distance between each speaker model and testing data, we use EMD which can calculate the distance between histograms with different bins. We conducted text-independent speaker identification experiments using the proposed method. Compared to results using the traditional GMM, the proposed method yielded relative error reductions of 32% for quantized data.

Neural networks have received considerable attention in sentence similarity measuring systems due to their efficiency in dealing with semantic composition. However, existing neural network methods are not sufficiently effective in capturing the most significant semantic information buried in an input. To address this problem, a novel weighted-pooling attention layer is proposed to retain the most remarkable attention vector. It has already been established that long short-term memory and a convolution neural network have a strong ability to accumulate enriched patterns of whole sentence semantic representation. First, a sentence representation is generated by employing a siamese structure based on bidirectional long short-term memory and a convolutional neural network. Subsequently, a weighted-pooling attention layer is applied to obtain an attention vector. Finally, the attention vector pair information is leveraged to calculate the score of sentence similarity. An amalgamation of both, bidirectional long short-term memory and a convolutional neural network has resulted in a model that enhances information extracting and learning capacity. Investigations show that the proposed method outperforms the state-of-the-art approaches to datasets for two tasks, namely semantic relatedness and Microsoft research paraphrase identification. The new model improves the learning capability and also boosts the similarity accuracy as well.

Previous studies of pattern recognition have shown that classifiers ensemble approaches can lead to better recognition results. In this paper, we apply the voting technique for the CoNLL-2010 shared task on detecting hedge cues and their scope in biomedical texts. Six machine learning-based systems are combined through three different voting schemes. We demonstrate the effectiveness of classifiers ensemble approaches and compare the performance of three different voting schemes for hedge cue and their scope detection. Experiments on the CoNLL-2010 evaluation data show that our best system achieves an F-score of 87.49% on hedge detection task and 60.87% on scope finding task respectively, which are significantly better than those of the previous systems.

It have been one open and significant topic for real-time applications to enhance the processing-speed of Boltzmann machines for long time. One effective way of solution of this problem is the augmentation of probability of neurons' state move. In this paper, a novel method, called a rejectionless method, was proposed and introduced into the Boltzmann machines for this augmentation. This method has a feature of independence on the ratio of neurons' state move. The efficiency of this method for speed-up was confirmed with the experiments of TSP and graph problem.

Protein-Protein Interaction Extraction (PPIE) from biomedical literatures is an important task in biomedical text mining and has achieved great success on public datasets. However, in real-world applications, the existing PPI extraction methods are limited to label effort. Therefore, transfer learning method is applied to reduce the cost of manual labeling. Current transfer learning methods suffer from negative transfer and lower performance. To tackle this problem, an improved TrAdaBoost algorithm is proposed, that is, relative distribution is introduced to initialize the weights of TrAdaBoost to overcome the negative transfer caused by domain differences. To make further improvement on the performance of transfer learning, an approach combining active learning with the improved TrAdaBoost is presented. The experimental results on publicly available PPI corpora show that our method outperforms TrAdaBoost and SVM when the labeled data is insufficient,and on document classification corpora, it also illustrates that the proposed approaches can achieve better performance than TrAdaBoost and TPTSVM in final, which verifies the effectiveness of our methods.

In this paper, we propose the construction of a web-based Question Answering (QA) system for restricted domain, which combines three resource information databases for the retrieval mechanism, including a Question&Answer database, a special domain documents database and the web resource retrieved by Google search engine. We describe a new retrieval technique of integrating a probabilistic technique based on OkapiBM25 and a semantic analysis which based on the ontology of HowNet knowledge base and a special domain HowNet created for the restricted domain. Furthermore, we provide a method of question expansion by computing word semantic similarity. The system is first developed for a middle-size domain of sightseeing information. The experiments proved the efficiency of our method for restricted domain and it is feasible to transfer to other domains expediently using the proposed method.

Biomedical event extraction is an important and challenging task in Information Extraction, which plays a key role for medicine research and disease prevention. Most of the existing event detection methods are based on shallow machine learning methods which mainly rely on domain knowledge and elaborately designed features. Another challenge is that some crucial information as well as the interactions among words or arguments may be ignored since most works treat words and sentences equally. Therefore, we employ a Bidirectional Long Short Term Memory (BLSTM) neural network for event extraction, which can skip handcrafted complex feature extraction. Furthermore, we propose a multi-level attention mechanism, including word level attention which determines the importance of words in a sentence, and the sentence level attention which determines the importance of relevant arguments. Finally, we train dependency word embeddings and add sentence vectors to enrich semantic information. The experimental results show that our model achieves an F-score of 59.61% on the commonly used dataset (MLEE) of biomedical event extraction, which outperforms other state-of-the-art methods.

Comparable Corpora are valuable resources for many NLP applications, and extensive research has been done on information mining based on comparable corpora in recent years. While there are not enough large-scale available public comparable corpora at present, this paper presents a bi-directional CLIR-based method for creating comparable corpora from two independent news collections in different languages. The original Chinese document collections and English documents collections are crawled from XinHuaNet respectively and formatted in a consistent manner. For each document from the two collections, the best query keywords are extracted to represent the essential content of the document, and then the keywords are translated into the language of the other collection. The translated queries are run against the collection in the same language to pick up the candidate documents in the other language and candidates are aligned based on their publication dates and the similarity scores. Results show that our approach significantly outperforms previous approaches to the construction of Chinese-English comparable corpora.

Collocation is a ubiquitous phenomenon in languages and accurate collocation recognition and extraction is of great significance to many natural language processing tasks. Collocations can be differentiated from simple bigram collocations to collocation frames (referring to distant multi-gram collocations). So far little focus is put on collocation frames. Oriented to translation and parsing, this study aims to recognize and extract the longest possible collocation frames from given sentences. We first extract bigram collocations with distributional semantics based method by introducing collocation patterns and integrating some state-of-the-art association measures. Based on bigram collocations extracted by the proposed method, we get the longest collocation frames according to recursive nature and linguistic rules of collocations. Compared with the baseline systems, the proposed method performs significantly better in bigram collocation extraction both in precision and recall. And in extracting collocation frames, the proposed method performs even better with the precision similar to its bigram collocation extraction results.

Compared to subword based Neural Machine Translation (NMT), character based NMT eschews linguistic-motivated segmentation which performs directly on the raw character sequence, following a more absolute end-to-end manner. This property is more fascinating for machine translation (MT) between Japanese and Chinese, both of which use consecutive logographic characters without explicit word boundaries. However, there is still one disadvantage which should be addressed, that is, character is a less meaning-bearing unit than the subword, which requires the character models to be capable of sense discrimination. Specifically, there are two types of sense ambiguities existing in the source and target language, separately. With the former, it has been partially solved by the deep encoder and several existing works. But with the later, interestingly, the ambiguity in the target side is rarely discussed. To address this problem, we propose two simple yet effective methods, including a non-parametric pre-clustering for sense induction and a joint model to perform sense discrimination and NMT training simultaneously. Extensive experiments on Japanese⟷Chinese MT show that our proposed methods consistently outperform the strong baselines, and verify the effectiveness of using sense-discriminated representation for character based NMT.

This paper attempts to establish a theory for a general auto-associative memory model. We start by defining a new concept called supporting function to replace the concept of energy function. As known, the energy function relies on the assumption of symmetric interconnection weights, which is used in the conventional Hopfield auto-associative memory, but not evidenced in any biological memories. We then formulate the information retrieving process as a dynamic system by making use of the supporting function and derive the attraction or asymptotic stability condition and the condition for convergence of an arbitrary state to a desired state. The latter represents a key condition for associative memory to have a capability of learning from variant samples. Finally, we develop an algorithm to learn the asymptotic stability condition and an algorithm to train the system to recover desired states from their variant samples. The latter called sample learning algorithm is the first of its kind ever been discovered for associative memories. Both recalling and learning processes are of finite convergence, a must-have feature for associative memories by analogy to normal human memory. The effectiveness of the recalling and learning algorithms is experimentally demonstrated.

In the European Telecommunication Standards Institute (ETSI), Distributed Speech Recognition (DSR) front-end, the distortion added due to feature compression on the front end side increases the variance flooring effect, which in turn increases the identification error rate. The penalty incurred in reducing the bit rate is the degradation in speaker recognition performance. In this paper, we present a nontraditional solution for the previously mentioned problem. To reduce the bit rate, a speech signal is segmented at the client, and the most effective phonemes (determined according to their type and frequency) for speaker recognition are selected and sent to the server. Speaker recognition occurs at the server. Applying this approach to YOHO corpus, we achieved an identification error rate (ER) of 0.05% using an average segment of 20.4% for a testing utterance in a speaker identification task. We also achieved an equal error rate (EER) of 0.42% using an average segment of 15.1% for a testing utterance in a speaker verification task.

Microblog data contains rich information of real-world events with great commercial values, so microblog-oriented natural language processing (NLP) tasks have grabbed considerable attention of researchers. However, the performance of microblog-oriented Chinese Word Segmentation (CWS) based on deep neural networks (DNNs) is still not satisfying. One critical reason is that the existing microblog-oriented training corpus is inadequate to train effective weight matrices for DNNs. In this paper, we propose a novel active learning method to extend the scale of the training corpus for DNNs. However, due to a large amount of partially overlapped sentences in the microblogs, it is difficult to select samples with high annotation values from raw microblogs during the active learning procedure. To select samples with higher annotation values, parameter λ is introduced to control the number of repeatedly selected samples. Meanwhile, various strategies are adopted to measure the overall annotation values of a sample during the active learning procedure. Experiments on the benchmark datasets of NLPCC 2015 show that our λ-active learning method outperforms the baseline system and the state-of-the-art method. Besides, the results also demonstrate that the performances of the DNNs trained on the extended corpus are significantly improved.

Chinese new words and their part-of-speech (POS) are particularly problematic in Chinese natural language processing. With the fast development of internet and information technology, it is impossible to get a complete system dictionary for Chinese natural language processing, as new words out of the basic system dictionary are always being created. A latent semi-CRF model, which combines the strengths of LDCRF (Latent-Dynamic Conditional Random Field) and semi-CRF, is proposed to detect the new words together with their POS synchronously regardless of the types of the new words from the Chinese text without being pre-segmented. Unlike the original semi-CRF, the LDCRF is applied to generate the candidate entities for training and testing the latent semi-CRF, which accelerates the training speed and decreases the computation cost. The complexity of the latent semi-CRF could be further adjusted by tuning the number of hidden variables in LDCRF and the number of the candidate entities from the Nbest outputs of the LDCRF. A new-words-generating framework is proposed for model training and testing, under which the definitions and distributions of the new words conform to the ones existing in real text. Specific features called "Global Fragment Information" for new word detection and POS tagging are adopted in the model training and testing. The experimental results show that the proposed method is capable of detecting even low frequency new words together with their POS tags. The proposed model is found to be performing competitively with the state-of-the-art models presented.