Search

With the economic development, loan business has rapidly developed in China. The risk that customers can’t repay their loans on time has increased. Therefore it is an important problem for financial organizations to approve the customers’ loan application or not. Typical machine learning methods for classification can be employed to mine customers’ financial information and give valuable judgments. However, these learning methods rely on shallow features, and the relationships between these features are not well studied. We investigate the function of Convolutional Neural Network (CNN) in this work, as it is successful in field of image recognition, speech recognition and natural language processing. We investigate four different CNN models. Experiments show that the fourth model with stochastic gradient descent algorithm and momentum achieves the best performance. Its accuracy and recall are 0.95 and 0.26 respectively.

Question classification (QC) is a basic task of question answering (QA) system. This task effectively narrows the range of candidate answers and improves the operating efficiency of the system by providing semantic restrictions for the subsequent steps of information retrieval and answer extraction. Due to the small number of words in the question, it is difficult to extract deep semantic information for the existing QC methods. In this work, we propose a QC method based on ERNIE and feature fusion. We approach this problem by first using ERNIE to generate word vectors, which we then use to input into the feature extraction model. Next, we propose to combine the hybrid neural network (CNN-BILSTM, which extracts features independently), highway network and DCU (Dilated Composition Units) module as the feature extraction model. Experimental results on Fudan university’s question classification data set and NLPCC(QA)-2018 data set show that our method can improve the accuracy, recall rate and F1 of the QC task.

With the rapid growth of text information, intelligent question answering has gained more attention than ever. In this paper, we focus on answer selection, one kind of question answering tasks. In this field, deep neural networks and attention mechanism have brought encouraging results. To improve the performance further, we investigate mixed embedding (word embedding and character embedding) representation for sentences to encode rich meaning. At the same time, we introduce a convolutional neural network (CNN) to compensate the loss of the max pooling layer in our attention based bidirectional Long Short-Term Memory (biLSTM) model. CNN features and the features from max pooling form final composite features, which are employed to select correct answers. Experimental results show that we can obviously improve the Mean Reciprocal Rank (MRR) performance by 6.0% with the help of mixed embedding and composite features. The MRR and ACC@1 score are 79.63% and 69.60% respectively.

The query error correction task is very important to improve user satisfaction and quality of query results. In traditional query error correction methods researchers mostly use a pipeline way to correct the error step by step. They rely heavily on manual annotation corpora. It is difficult to take into account the global effect. In this paper, we present a character-based end-to-end Sequence to Sequence (Seq2Seq) method with attention mechanism. It also incorporates the neural network language model trained on unlabeled corpora to solve the task of query correction. It can unify the modeling of different error types in query error correction and effectively overcome the shortcomings of traditional methods in query error correction tasks. Experiments show that this method can effectively capture the long-distance knowledge to correct errors, and through the Simple Recurrent Unit (SRU) it can be as good as Long Short-Term Memory (LSTM). However, there has been a significant improvement in processing time. This point is very important in query error correction tasks.

Chinese Spelling Check (CSC) is very important for Chinese language processing. To utilize supervised learning for CSC, one of the main challenges is that high-quality annotated corpora are not enough in building models. This paper proposes new approaches to automatically build the corpora of CSC based on the input method. We build two corpora: one is used to check the errors in the texts generated by the Pinyin input method, called p-corpus, and the other is used to check the errors in the texts generated by the voice input method, called v-corpus. The p-corpus is constructed using two methods, one is based on the conversion between Chinese characters and the sounds of the characters, and the other is based on Automatic Speech Recognition (ASR). The v-corpus is constructed based on ASR. We use the misspelled sentences in real language situation as the test set. Experimental results demonstrate that our corpora can get a better checking effect than the benchmark corpus.

The sentimental analysis of Chinese microblog is a crucial task for social network related applications, such as internet marketing, public opinion monitoring, etc. This paper proposes a two-phase computing model for microblog sentimental analysis, including topic classification for posts and topic sentimental analysis respectively. In the first phase, the Latent Dirichlet Allocation(LDA) model is employed into our model for topic classification. The topics of posts are scattered into the microblogs, it has some uncertainty. The LDA model can classify the fuzzy topics. In the second phase, sentimental dictionary and emotion knowledge are performed into our model for topic sentimental analysis. HowNet as the sentimental dictionary is used to the sentimental tendency analysis. The emotion knowledge mainly uses symbols in microblog. Besides of sentimental knowledge, the sliding window for sentimental analysis also introduced into the model as the modification at the sentence level. The experimental results show that this method achieves good performance.

In this paper, we discussed the relation between the attributes of a concept and its similarity position with other concepts. We constructed a function to map a similarity position of a concept to its coordinates in a geometry space. The coordinates can be further mapped to the attributes through another function. We constructed the functions by distance geography methods and proved that such functions do exist under some conditions. This work will benefit attribute retrieval tasks.