Research Of Chinese Anaphora Resolution And Its Application In Vision Object Detection

With the rapid development of computer and Internet technology, different kinds of information have been growing exponentially. In the face of the huge amount of heterogeneous information, it is difficult for us to get what we want. How to correctly extract, process and classify the information, as well as make it easy to retrieve, has become a hot yet difficult problem of natural language processing. Anaphora Resolution is one of the key tasks to solve the above problem. Therefore, the research on Anaphora Resolution gets more and more attention. Anaphora Resolution can not only eliminate the ambiguity in the text, but also relate the different descriptions that pointing to the same entity from different signal sources. Therefore, anaphora resolution is widely used in many fields, such as natural language processing, database and machine vision, etc..This dissertation focuses on the Chinese anaphora resolution methods and its application in the field of vision object detection, and the contents of this dissertation can be summarized as following:(1) Six important conclusions that affect the performance of Chinese anaphora resolution models are derived by comparison experiments.Six classic Chinese anaphora resolution models are compared on ACE2005 and OntoNotes5.0 based on the same platform, same corpus and same features. The factors affecting the accuracy of Chinese anaphora resolution models are analyzed and discussed. Furthermore, the models that are worthy of further research and exploration under the condition of lacking corpus are pointed out, based on which, the following work of this dissertation are expanded.(2) A metric-optimized Laplacian SVM is proposed for Chinese anaphora resolution method.Compared to English corpus, Chinese corpus is much smaller, and the unlabeled samples are much more than the labeled ones. To efficiently explore the similarity and correlations between labeled and unlabeled samples for deriving more accurate classification model, a data-driven based method is proposed to learn the optimal distance metric for Laplacian SVM. The proposed method takes similarity constraints between sample-pairs into consideration so that the similarities of in-class samples are higher than those of between-class samples. Moreover, the Fisher discrimination criterion is also introduced, which can help to make the samples from the same class have small scatters, wheres the samples from different classes have large scatters in the new metric space. Furthermore, the proposed linear metric-optimized method is generalize to its kernel form, which helps the Laplacian SVM make use of the kernel function for non-linear classification. Compared with the supervised methods and other four semi-supervised methods on the ACE2005 and OntoNotes5.0 Chinese corpus, the proposed method, both the linear form and kernel form, achieves the comparative or best performance, while uses fewer labeled samples.(3) A modified multi-pass sieve model is proposed for Chinese anaphora resolution.The model first adds a new semantic-based sieve to the original model for incorporating word sense information, which is used to solve resource constraints by importing the Web semantic knowledge. Second, focusing on the fact of low number type recognition rate, the sieve for number type recognition in the original model is modified to improve the pronoun resolution accuracy. Finally, the mention detection sieve is modified based on the Chinese character. The proposed model is evaluated on five different testing methods on the ACE2005 corpus, and the test results show that the proposed model outperforms four other baseline models.(4) A method to combine textual and visual information for the vision object detection is proposed.Text and image are two of the most important sources that human beings perceive the world. Focusing on the flaws of the method of typical object detection problem-pedestrain detection problem, the dissertion inroduce textual information to improve the detection accuracy of pedestrian object detection in image on the basis of traditional vision-based methods. The coreference relation between object candidates in images and mentions in text are inferred to combine textual and image information. First, the candidate visual objects are localized by using a vision-based method; second, the text mentions corresponded to visual objects are obtained by a textual analysis method. Then, a Markov Random Field-based model is proposed to infer the coreference relation between the candidate visual objects and textual mentions, and thus the detection accuracy can be improved by combining the textual information. The effectiveness of the proposed method is demonstrated on the Caltech Pedestrian Detection Benchmark, which is enriched with lingual descriptions. It is shown that the proposed method significantly outperform the state-of-the-art visual-based pedestrian detection methods, and is able to reliably estimate the text-to-image coreference. Furthermore, by using textual and visual information, it is successful to deal with anaphora resolution in text, improving upon the baseline anaphora resolution model by 4%.