Contextual advertising focuses on placing suitable advertisements on web pages. To attract user’s intention, the advertisements should be highly related to the target web page. The most effective way to do contextual advertising is ontology-based matching algorithms. The main problem of such algorithms is the difficulty in constructing and populating the ontology for matching advertisements. In this paper, we propose an automatic construction method for advertisement ontology. The construction method searches related documents from Web, extracts keywords and weights keywords for concepts. The weighted keywords are treated as instances of concepts and used to generate centroid vectors for concepts. In order to weight keywords in a proper way, we raise a formula WebSSR (Super-Subordinate Relation by Web). WebSSR weights words based on the probabilities that they have Specific Relations with the target concept. We compare our formula with LDA, NGD, WebJaccard, WebOverlap, WebDice and WebPMI, and our formula outperforms all of them. Experiment results also show that our method is more effective than five baseline methods: Bayesian, SVM, SLSA, LDA and Paragraph2Vec.

Recently, neural sequence-to-sequence models have made impressive progress in abstractive document summarization. Unfortunately, as neural abstractive summarization research is in a primitive stage, the performance of these models is still far from ideal. In this paper, we propose a novel method called Neural Abstractive Summarization with Diverse Decoding (NASDD). This method augments the standard attentional sequence-to-sequence model in two aspects. First, we introduce a diversity-promoting beam search approach in the decoding process, which alleviates the serious diversity issue caused by standard beam search and hence increases the possibility of generating summary sequences that are more informative. Second, we creatively utilize the attention mechanism combined with the key information of the input document as an estimation of the salient information coverage, which aids in finding the optimal summary sequence. We carry out the experimental evaluation with state-of-the-art methods on the CNN /Daily Mail summarization dataset, and the results demonstrate the superiority of our proposed method.

Deep learning shows its superiority in many domains such as computing vision, nature language processing, and speech recognition. In music recommendation, most deep learning-based methods focus on learning users’ temporal preferences using their listening histories. The cold start problem is not addressed, however, and the music characteristics are not fully exploited by these methods. In addition, the music characteristics and the users’ temporal preferences are not combined naturally, which cause the relatively low performance of music recommendation. To address these issues, we proposed a Deep Temporal Neural Music Recommendation model (DTNMR) based on music characteristics and the users’ temporal preferences. We encoded the music metadata into one-hot vectors and utilized the Deep Neural Network to project the music vectors to low-dimensional space and obtain the music characteristics. In addition, Long Short-Term Memory (LSTM) neural networks are utilized to learn about users’ long-term and short-term preferences from their listening histories. DTNMR alleviates the cold start problem in the item side using the music medadata and discovers new users’ preferences immediately after they listen to music. The experimental results show DTNMR outperforms seven baseline methods in terms of recall, precision, f-measure, MAP, user coverage and AUC.

Automatic question generation from text or paragraph is a great challenging task which attracts broad attention in natural language processing. Because of the verbose texts and fragile ranking methods, the quality of top generated questions is poor. In this paper, we present a novel framework Automatic Chinese Question Generation (ACQG) to generate questions from text or paragraph. In ACQG, we use an adopted TextRank to extract key sentences and a template-based method to construct questions from key sentences. Then a multi-feature neural network model is built for ranking to obtain the top questions. The automatic evaluation result reveals that the proposed framework outperforms the state-of-the-art systems in terms of perplexity. In human evaluation, questions generated by ACQG rate a higher score.

With the development of online advertisements, clickbait spread wider and wider. Clickbait dissatisfies users because the article content does not match their expectation. Thus, clickbait detection has attracted more and more attention recently. Traditional clickbait-detection methods rely on heavy feature engineering and fail to distinguish clickbait from normal headlines precisely because of the limited information in headlines. A convolutional neural network is useful for clickbait detection, since it utilizes pretrained Word2Vec to understand the headlines semantically, and employs different kernels to find various characteristics of the headlines. However, different types of articles tend to use different ways to draw users’ attention, and a pretrained Word2Vec model cannot distinguish these different ways. To address this issue, we propose a clickbait convolutional neural network (CBCNN) to consider not only the overall characteristics but also specific characteristics from different article types. Our experimental results show that our method outperforms traditional clickbait-detection algorithms and the TextCNN model in terms of precision, recall and accuracy.

With the development of large-scale knowledge bases (KBs), knowledge-based question answering (KBQA) has become an important research topic in recent years. The key task in KBQA is relation detection, which is the process of finding a compatible answer type for a natural language question and generating its corresponding structured query over a KB. However, existing systems often rely on shallow probabilistic methods, which are less expressive than deep semantic representation methods. In addition, since KBs are still far from complete, it is necessary to develop a new strategy that leverages unstructured resources outside of KBs. In this work, we propose a novel Question Answering method with Relation Detection and Textual Evidence (QARDTE). First, to address the semantic gap problem in relation detection, we use bidirectional long-short term memory networks with different levels of abstraction to better capture sentence structures. Our model achieves improved results with robustness against a wide diversity of expressions and questions with multiple relations. Moreover, to help compensate for the incompleteness of KBs, we utilize external unstructured text to extract additional supporting evidence and combine this evidence with relation information during the answer re-ranking process. In experiments on two well-known benchmarks, our system achieves F1 values of 0.558 (+2.8%) and 0.663 (+5.7%), which are state-of-the-art results that show significant improvement over existing KBQA systems.

With the development of information technology, it becomes more and more difficult to retrieve information from the internet for users. Question Answering (QA) is one of the methods to solve this problem. The users type natural language questions and get answers in QA systems. However, most QA systems only return a word or several words to the user, which is not friendly enough. The users are more willing to receive not only answers but also additional introductions or reasons. In this work, we propose a Nature Language Question Answering system which utilizes Seq2Seq model and Generative Adversarial Network (GAN) to generate answers with more information for users. To our best knowledge, this is the first work generating natural language answers in Question Answering domain. Our experiment results show NLQA can generate readable answers for users.

In recommender systems (RS), many models are designed to predict ratings of items for the target user. To improve the performance for rating prediction, some studies have introduced tags into recommender systems. Tags benefit RS considerably, however, they are also redundant and ambiguous. In this paper, we propose a hybrid deep learning model TRSDL (tag-aware recommender system based on deep learning) to improve the performance of tag-aware recommender systems (TRS). First, TRSDL uses pre-trained word embeddings to represent user-defined tags, and constructs item and user profiles based on the items’ tags set and users’ tagging behaviors. Then, it utilizes deep neural networks (DNNs) and recurrent neural networks (RNNs) to extract the latent features of items and users, respectively. Finally, it predicts ratings from these latent features. The model not only addresses tag limitations and takes advantage of semantic tag information but also learns more advanced implicit features via deep structures. We evaluated our proposed approach and several baselines on MovieLens-20 m, and the experimental results demonstrate that TRSDL significantly outperforms all the baselines (including the state-of-the-art models BiasedMF and I-AutoRec). In addition, we also explore the impacts of network depth and type on model performance.

Automatically generation of an image description is a challenging task which attracts broad attention in artificial intelligence. Inspired by methods of computer vision and natural language processing, different approaches have been proposed to solve the problem. However, captions generated by the existing approaches have been lack of enough contextual information to describe the corresponding images completely. The labeled captions in the training set only basically describe images and lack of enough contextual annotations. In this paper, we propose a Weakly-supervised Image Captioning Approach (WICA) to generate captions containing rich contextual information, without complete annotations for the contextual information in datasets. We utilize encoder-decoder neural networks to extract basic captioning features and leverage object detection networks to identify contextual features. Then, we encode the two levels of features by a phrase-based language model in order to generate captions with rich contextual information. The comprehensive experimental results reveal that proposed model outperforms the existing baselines in terms of on the richness and reasonability of contextual information for image captioning.

In the mobile Internet era, users access interesting information in a continuous manner rather than as one-time results through search engines. The traditional link-based ranking algorithms typically return the relevant "popular'' web pages. The current, most important web pages are ranked lower than these pages. Furthermore, most of the results are repeated when the user submits the same query days later. In this paper, we have described a novel service called tracking engine. The tracking engine allows users to enter and save queries, displays time-sensitive information, and notifies users when new, relevant information appears. To the best of our knowledge, this is the first solution seen in such a service in the mobile Internet era. First, our tracking engine called Tianji crawls the web pages based on time priority and constructs a new index structure, which enables a faster match of web pages to related keywords. Then, we develop a ranking model based on the correlation between time and importance. The experimental results show that the ranking model of Tianji has better performance than existing time-sensitive ranking methods in terms of timeliness and relevance.

In the web advertising domain, contextual advertising and sponsored search are two of the main advertising channels used to display related advertisements on web pages. A major challenge for contextual advertising is to match advertisements and web pages based on their semantics. When a web page and its semantically related advertisements contain many different words, the performance of the traditional methods can be very poor. In particular, there are few studies presented for Chinese contextual advertising that are based on semantics. To address these issues, we propose an ontology-based approach to Chinese semantic advertising. We utilize an ontology called the Taobao Ontology and populate it by automatically adding related phrases as instances. The ontology is used to match web pages and advertisements on a conceptual level. Based on the Taobao Ontology, the proposed method exploits seven distance functions to measure the similarities between concepts and web pages or advertisements. Then, the similarities between web pages and advertisements are calculated by considering the ontology-based similarities as well as term-based similarities. The empirical experiments indicate that our method is able to match Chinese web pages and advertisements with a relatively high accuracy. Among the seven distance functions, Cosine distance and Tanimoto distance show the best performance in terms of precision, recall, and F-measure. In addition, our method outperforms two contextual advertising methods, i.e., the impedance coupling method and the SVM-based method.

Semantic Advertising is a new kind of web advertising to find the most related advertisements for web pages semantically. In this way, users are more likely to be interest in the related advertisements when browsing the web pages. A big challenge for semantic advertising is to match advertisements and web pages in a conceptual level. Especially, there are few studies proposed for Chinese semantic advertising. To address this issue, we proposed an adaptive method to construct an ontology automatically for matching Chinese advertisements and web pages semantically. Seven distance functions are exploited to measure the similarity between advertisements and web pages. Based on the empirical experiments, we found the proposed method shows a promising result in terms of precision, and among the distance functions, the Tanimoto distance function outperforms the other six distance functions.