Cross-media analysis and reasoning: advances and directions
- 2.1k Downloads
Cross-media analysis and reasoning is an active research area in computer science, and a promising direction for artificial intelligence. However, to the best of our knowledge, no existing work has summarized the state-of-the-art methods for cross-media analysis and reasoning or presented advances, challenges, and future directions for the field. To address these issues, we provide an overview as follows: (1) theory and model for cross-media uniform representation; (2) cross-media correlation understanding and deep mining; (3) cross-media knowledge graph construction and learning methodologies; (4) cross-media knowledge evolution and reasoning; (5) cross-media description and generation; (6) cross-media intelligent engines; and (7) cross-media intelligent applications. By presenting approaches, advances, and future directions in cross-media analysis and reasoning, our goal is not only to draw more attention to the state-of-the-art advances in the field, but also to provide technical insights by discussing the challenges and research directions in these areas.
Key wordsCross-media analysis Cross-media reasoning Cross-media applications
Along with the progress of human civilization and the development of science and technology, information acquisition, transmission, processing, and analysis have gradually changed from one form of media to multiple types of media such as text, image, video, audio, and stereo picture. Different media types on various platforms and modalities from social, cyber, and physical spaces are now mixed together to demonstrate rich natural and social properties. As a whole they represent comprehensive knowledge and reflect the behavior of individuals and groups. Consequently, a new form of information is recognized, known as cross-media information.
Over the past several decades, as the requirements for data management and utilization have increased significantly, multimedia information processing and analysis has been a research hotspot (Lew et al., 2006). However, previous studies were devoted mainly to scenarios involving a single media. Research in cognitive science indicates that in the human brain, cognition of the environment is through the fusion of multiple sensory organs (McGurk and MacDonald, 1976). Although the representations of different media types are heterogeneous, they may share the same semantics, and have rich latent correlations. Consider the topic of ‘bird’ as an example. All of the texts, images, videos, audio clips, and stereo pictures about this topic describe the same semantic concept ‘bird’ from complementary aspects. As a result, due to limitations in information diversity, traditional single-media analysis methods have difficulty in achieving the goal of semantic extraction from multiple modalities, and cannot deal with the analysis of cross-media data. Meanwhile, traditional reasoning methods are mainly text-based and perform reasoning under fully defined premises. They cannot deal with cross-media scenarios with sophisticated compositions, different representations, and complex correlations. Therefore, a key problem in research and application has been how to simulate the human brain’s process of transforming environmental information to analytical models through vision, audition, language, and other sensory channels, and further to realize cross-media analysis and reasoning.
The topic of cross-media analysis and reasoning has attracted considerable research interest. With respect to cross-media analysis, existing studies focus mainly on modeling correlations and generating a uniform representation of two media types as in the popular correlation analysis method, called canonical correlation analysis (CCA) (Hotelling, 1936). Though there are limited studies on cross-media reasoning so far, it is an important future direction to extend traditional text-based reasoning methods to cross-media scenarios. There are also wide prospects for applications in cross-media analysis and reasoning. Effective yet efficient cross-media methods can provide more flexible and convenient ways to retrieve and manage multimedia big data. Users would like to adopt the cross-media intelligent engine for applications such as cross-media retrieval, and cross-media technology is also useful for important application scenarios, such as web content monitoring, web information trend analysis, and healthcare data fusion and reasoning. However, there still exist important challenges for cross-media intelligent applications.
Cross-media analysis and reasoning has been an active research area in computer science, and an important future direction in artificial intelligence. As discussed in Pan (2016), cross-media intelligence plays the role of a cornerstone in artificial intelligence, through which the machines can recognize the external environment. Although considerable improvement has been made in the research of cross-media analysis and reasoning (Rasiwasia et al., 2010; Yang et al., 2012; Peng et al., 2016a; 2016b), there remain some important challenges and unclear points in future research directions. In this paper, we give a comprehensive overview of not only the advances achieved by existing studies, but also future directions for cross-media analysis and reasoning. The aim is to attract more researchers to the research field in cross-media analysis and reasoning, and thus we provide insights by discussing challenges and research directions, to facilitate new studies and applications on this new and exciting research topic.
2 Cross-media analysis and reasoning
The advances and directions in cross-media analysis and reasoning can be summarized as seven parts: (1) theory and model for cross-media uniform representation; (2) cross-media correlation understanding and deep mining; (3) cross-media knowledge graph construction and learning methodologies; (4) cross-media knowledge evolution and reasoning; (5) cross-media description and generation; (6) cross-media intelligent engines; (7) cross-media intelligent applications. In this section, we will provide descriptions of these seven parts, so as to present a comprehensive overview of cross-media analysis and reasoning.
2.1 Theory and model for cross-media uniform representation
To the best of our knowledge, the first well-known cross-media model is based on CCA (Rasiwasia et al., 2010). It learns a commonly shared space by maximizing the correlation between pairwise co-occurring heterogeneous data and performs projection by linear functions. Although the scheme is simple, it has inspired subsequent studies. CCA has many variants (Andrew et al., 2013; Gong et al., 2014; Rasiwasia et al., 2014). For example, Andrew et al. (2013) extended this method using a deep learning technique to learn the correlations more comprehensively than those using CCA and kernel CCA. These methods can, for the most part, model only the correlations of two media types. To overcome this limitation, researchers have also attempted to develop datasets and methods for scenarios with more media types. For example, the newly constructed XMedia dataset (http://www.icst.pku.edu.cn/mipl/XMedia) is the first dataset containing five media types (text, image, video, audio, and 3D model), and methods such as those proposed by Zhai et al. (2014) and Peng et al. (2016b) can jointly model the correlations and semantic information in a unified framework with graph regularization for the five media types on the XMedia dataset. Yang et al. (2008) introduced another model called the multimedia document (MMD) to represent data, where each MMD is a set of media objects of different modalities but carrying the same semantics. The distances between MMDs are related to each modality, and in this way we can perform cross-media retrieval. Daras et al. (2012) employed a radial basis function (RBF) network to address the problem of missing modalities. However, the main problem with the MMD is that it only handles data from different modalities together, which is not flexible in many applications. Most cross-media representation learning models still belong to subspace learning techniques.
The topic model is another frequently used technique in cross-media uniform representation learning tasks, assuming that heterogeneous data containing the same semantics shares some latent topics. For example, Roller and Schulte im Walde (2013) integrated visual features into latent Dirichlet allocation (LDA) and proposed a multimodal LDA model to learn representations for textual and visual data. Wang Y et al. (2014) proposed a scheme called the multimodal mutual topic reinforce model (M3R), which seeks to discover mutually consistent semantic topics via appropriate interactions between model factors. These schemes represent data as topic distributions, and similarities are measured by the likelihood of observed data in terms of latent topics. Metric learning is usually performed if we know which data pairs are similar and which are dissimilar from heterogeneous modalities. An appropriate distance metric is designed to measure heterogeneous similarity, and learned using the given labeled data pairs to achieve the best performance. When the learned metric is decomposed into modality-specific projection functions (Wu et al., 2010), data can be explicitly projected into a uniform representation as CCA does. Apart from the above-mentioned models, Mao et al. (2013) proposed a manifold-based model called parallel field alignment retrieval (PFAR), which considers cross-media retrieval as a manifold alignment problem using parallel fields.
Although there are significant research efforts on uniform representation learning for cross-media analysis tasks, a large gap still exists between these methods and user expectations. This is caused by the fact that existing schemes still have not achieved a satisfactory performance; i.e., their accuracies are far from acceptable. Therefore, we still need to investigate better uniform representation methods for cross-media research.
2.2 Cross-media correlation understanding and deep mining
Cross-media correlations describe specific types of statistical dependencies among homogeneous and heterogeneous data objects. For example, if two images are taken from the same location, they may be intrinsically correlated from content, attribute, and topic perspectives, and thus they may share certain levels of intrinsic semantic consistency. The content in the paragraphs and social comments on a video webpage is semantically related to the content of the video itself. The aim of cross-media correlation learning is to construct metrics on heterogeneous data representation to measure how they are semantically relevant.
Following another line of research, researchers from the database community have investigated the correlations and fusion among unstructured, semi-structured, and structured data. However, most of these studies are based on low-level features and formats. Few studies are focusing on multimodal content and high-level correlations, e.g., generating a description for the entities by fusing semi-structured Wiki data and unstructured web data. Moreover, cross-media data is not only from different modalities and structures, but also from different sources. The study of associating and fusing cross-media data from different sources remains in its infancy, e.g., objective data and subjective user-generated content (UGC), user data from different online social networks (OSNs), and cross-space data from cyber and physical spaces.
In cross-media deep mining, the knowledge base is manually and professionally edited by experts in traditional expert systems. Currently, many studies are focusing on extracting and learning knowledge from data automatically, e.g., Google Knowledge Vault (Dong et al., 2014). However, similar to data, knowledge is essentially cross-media. Recently we have seen a rapid development of different types of intelligent perceptions, e.g., vision-based environmental perception in Visual SLAM (Fuentes-Pacheco et al., 2015) and multimodal based human-computer interaction in gesture and action recognition (Rautaray and Agrawal, 2015). Moreover, ubiquitous perception has received increasing attention these days (Adib et al., 2015). Development in the above areas provides opportunities to research the problem of cross-media knowledge mining. While critical challenges exist in constructing the cross-media knowledge base, it is of great theoretical and technical significance to combine perceptions from different modalities to supplement and improve the current text-based knowledge base.
Despite the achievements in cross-media correlation understanding, there is still a long way to go in this research direction. Basically, existing studies construct correlation learning on cross-media data with representation learning, metric learning, and matrix factorization, which are usually performed in a batch learning fashion and can capture only the first-order correlations among data objects. How to develop more effective learning mechanisms to capture the high-order correlations and adapt to the evolution that naturally exists among heterogeneous entities and heterogeneous relations, is the key research issue for future studies in cross-media correlation understanding.
2.3 Cross-media knowledge graph construction and learning methodologies
The aim of cross-media knowledge graph construction is to represent framed rules, values, experiences, contexts, instincts, and insights with entities and relations from general to specific domains (Davenport and Prusak, 1998). In cross-media research, the entities and relations are defined and extracted from not only the textual data corpus, but also numerous loosely correlated data modalities including texts, images, videos, and other related information sources. Cross-media knowledge graphs provide essential computable knowledge representation structures for semantic correlation analysis and cognition-level reasoning in cross-media context, facilitating theoretical and technical development in cross-media intelligence and a diversified range of applications.
The second area of focus on knowledge graphs is how to deploy knowledge graphs to enhance the performance and user experience in information retrieval and web applications, especially in the era of big data. As a pioneering work, Garfield (2004) developed the HistCite software to generate knowledge graphs in academic literature, which led to the birth of the academic search engine CiteSeer. The Knowledge Graph released by Google in 2012 (Singhal, 2012) provided a next-generation information retrieval service with ontology-based intelligent search based on free-style user queries. Similar techniques, e.g., Safari, were developed based on achievements in entity-centric search (Lin et al., 2012). However, existing entity-based search engines cannot perform fully automatic content parsing on heterogeneous modalities, and thus they cannot provide entity-based information retrieval for cross-media content.
To transform the web of data into a web of knowledge (Suchanek and Weikum, 2014), several issues should be considered in research on cross-media knowledge graphs. First of all, effective and efficient techniques for entity extraction and relation construction from heterogeneous cross-media information sources should be studied. Second, information search and retrieval based on cross-media knowledge graphs should be investigated to provide more effective knowledge harvesting and information seeking mechanisms for more diverse application contexts. Third, mining and reasoning in cross-media knowledge graphs should be developed to facilitate knowledge acquisition and high-level reasoning for real applications. Finally, knowledge-driven cross-media learning models will be required in the near future to achieve more generalization and learning capabilities, resulting in more advanced cross-media intelligence.
2.4 Cross-media knowledge evolution and reasoning
In addition, it has been shown that some learning mechanisms, such as reinforcement learning and transfer learning, can be helpful for constructing more complex intelligent reasoning systems (Lazaric, 2012). Furthermore, lifelong learning (Lazer et al., 2014) is the key capability of advanced intelligence systems. For example, Google DeepMind has constructed a machine intelligence system based on a reinforcement learning algorithm (Gibney, 2015), which beat humans at classic video games. Recently, AlphaGo, developed by Google DeepMind, has been the first computer Go program that can beat a top professional human Go player. It even beat the world champion Lee Sedol in a five-game match. We have witnessed increasing numbers of intelligence systems winning human-machine competitions.
However, the knowledge and reasoning process in the real world usually involves collaboration among language, vision, and other types of media data. Most existing intelligent systems exploit only the information from a single media type, such as text, to perform reasoning processes. There have been some recent works involving reasoning on cross-media data. Visual question answering (VQA) can be regarded as a good example of cross-media reasoning (Antol et al., 2015). VQA aims to provide natural language answers for questions given in the form of combination of the image and natural language. Johnson et al. (2015) attempted to improve the accuracy of image retrieval with the assistance of the scene graph, which also shows the idea of cross-media reasoning. A scene graph presents objects and their attributes and relationships, which can be used to guide image retrieval at the semantic level. However, it is still hard for these systems to make full use of the rich semantic information contained in complementary media types, and they cannot perform complex cross-media analysis and reasoning on multimedia big data. Therefore, the problem of performing cross-media reasoning based on multiple media types rather than on only text information, has become important in both research and application areas. Note that there is little research on cross-media knowledge evolution and reasoning, and many key problems need to be solved, which include, for instance, the acquisition, representation, mining, learning, and reasoning of cross-media knowledge, and the construction of large-scale cross-media knowledge bases. We still need to confront the significant challenges that are involved in constructing cross-media reasoning systems for real applications.
To address the problems noted above, several issues should be studied further. First, it is important to study data-driven and knowledge-guided cross-media knowledge learning methods. Second, cross-media reasoning frameworks based on semantic understanding should be constructed with technologies such as cross-media deep learning and multi-instance learning. Third, never-ending knowledge acquisition, mining, and evolution processes should be comprehensively investigated in future work.
2.5 Cross-media description and generation
Existing studies on visual content description can be divided into three groups. The first group, based on language generation, first understands images in terms of objects, attributes, scene types, and their correlations, and then connects these semantic understanding outputs to generate a sentence description using natural language generation techniques, e.g., templates (Yang et al., 2011), n-grams (Kulkarni et al., 2011), and grammar rules (Kuznetsova et al., 2014). These methods are direct and intuitive, but the sentences generated are limited by their syntactic dependency and thus are inflexible.
The second group covers retrieval-based methods, retrieving content that is similar to a query and transferring the descriptions of the similar set to the query. According to the differences in the retrieval feature space, studies in this group include two types, i.e., retrieval in a uni-modal space (Ordonez et al., 2011) and in a multimodal space (Hodosh et al., 2013). The former aims to search for similar images or videos in the visual feature space, and the latter projects images or videos and sentence features into a common multimodal space, and searches for similar content in the projected space. Sentences obtained with these methods are more natural and grammatically correct, but they usually suffer with regard to generating variable-length and novel sentences.
The third group is based on deep neural networks, employing the CNN-RNN codec framework, where the convolutional neural network (CNN) is used to extract features from images, and the recursive neural network (RNN) (Socher et al., 2011) or its variant, the long short-term memory network (LSTM) (Hochreiter and Schmidhuber, 1997), is used to encode and decode language models. These methods typically use neural networks for both image-text embedding and sentence generation (Karpathy and Li, 2015; Vinyals et al., 2015), and visual attention (Xu et al., 2015) or semantic guidance (Jia et al., 2015) is also integrated in the model learning to further improve the performance. Compared with the other methods, the deep models benefit from a stronger feature expression ability from CNN and capture dynamic spatiotemporal information with RNN, and thus they receive more attention. However, it is still a preliminary exploration and there exist many problems regarding further research: (1) As the parameter size of deep neural network is huge, it demands large amounts of annotated data for training and is easy to overfit, which makes sentence generation depend heavily on the training set; (2) The global features from CNN have difficulty in representing local objects accurately, which results in incorrect or missing descriptions of local objects, especially their correlation in images.
In conclusion, the current research is centered mainly on natural language descriptions of singlemedia content, and improvements are needed in the areas of training set collection and application, model building, and efficient learning and optimization modeling with human cognition. Furthermore, the cross-media descriptions of text, image, video, and audio are rarely involved, such as image generation from text and video generation from audio. Considering that human cognition is an integrated understanding procedure of different types of sensory information, it becomes a very challenging but valuable task to implement a comprehensive and accurate description of multimodal information with natural language processing. The connections with complex cognition, human emotion, and logical reasoning are also attractive areas for in-depth exploration.
2.6 Cross-media intelligent engines
The intelligent engine is a kind of intelligent analysis and reasoning system having specific purposes and common knowledge. With the rapid developments in artificial intelligence, some international companies and research institutions have implemented text-based artificial intelligent systems with specific capabilities. Technology companies such as Google, Baidu, and Microsoft have proposed the concept of intelligent search and the framework for search techniques (Uyar and Aliyu, 2015). Based on the highly effective indexing of big data, intelligent search attempts to realize intelligent and humanized information services, allowing users to retrieve whatever they want with input in natural language forms. It can provide more convenient and accurate search results than traditional search engines. In the field of medical treatment, researchers have also proposed the technological concept of the intelligent medical search engine (Luo and Tang, 2008).
However, cross-media big data is naturally multimodal and cross-domain, employing sophisticated compositions, different representations, and complex correlations. Existing intelligent systems and frameworks depend heavily on the structured input and knowledge of specific domains. They cannot adapt to the characteristics of cross-media data, and cannot cope with the increasingly complex needs of general tasks (such as information retrieval) and specific tasks (such as content monitoring) in cross-media scenarios, which makes it very hard for them to realize cross-media intelligent analysis and reasoning. To address these problems, it is essential to develop an efficient cross-media intelligent engine with abilities in autonomic learning and evolution. The efficient intelligent engine would act as a bridge between technologies and applications, which could integrate cross-media uniform representation, correlation learning, knowledge evolution, reasoning, and so on. Such an engine would provide cross-media analysis and reasoning services, and be a computing platform for cross-media intelligent applications.
2.7 Cross-media intelligent applications
The advent of the artificial intelligence era and the availability of huge amounts of cross-media data have been revolutionizing the landscape in all industry sectors. Among these, cross-media web content monitoring, web information trend analysis, and healthcare data fusion and reasoning are three key applications, which if well addressed would present important models and demonstration significance to all other areas. We will briefly review the preliminary background, previous studies, as well as the existing challenges to be confronted.
iMonitor: The Internet is recognized as one of the most influential factors for the stability of human society. Many countries have built intelligent systems to monitor the content propagating or streaming over the Internet, such as the PRISM system in the US, the Tempora system in the UK, and the SORM system in Russia. At the same time, China is developing a set of web content monitoring systems, such as the Golden Shield Project for the Ministry of Public Security of China. However, existing monitoring systems work mainly in the form of passive sampling-post hoc analysis, which limits the usefulness of existing systems, and raises three challenges in the intelligent systems community, namely (1) time lag, (2) insufficient coverage, and (3) high cost, especially considering the diversity of cross-media data.
Many IT giants have joined the healthcare analytics community; e.g., IBM released Watson Healthcare (http://spectrum.ieee.org/computing/software/), Google announced DeepMind (https://deepmind.com/health), and Baidu just released Baidu Medical Brain. In spite of their usefulness in certain areas, the applicability of existing models and algorithms (Kumar et al., 2012; Chen Y et al., 2013; Yuan et al., 2014) is limited due to (1) inability to perform cross-media fusion and analysis (Chen et al., 2007), (2) lack of supervision from domain experts (Chen Y et al., 2013), and (3) poor adaptability toward different medical paradigms.
In this paper, we have presented an overview of cross-media analysis and reasoning. The advances achieved by existing studies, as well as the major challenges and open issues, have been shown in the overview. From the seven parts of this paper, it can be seen that cross-media analysis and reasoning has been a key problem of research, and has wide prospects for application. The introduction and discussion in this paper are expected to attract more research interest to this area, and provide insights for researchers on the relevant topics, so as to inspire future research in cross-media analysis and reasoning.
The authors would like to thank Peng CUI, Shi-kui WEI, Ji-tao SANG, Shu-hui WANG, Jing LIU, and Bu-yue QIAN for their valuable discussions and assistance.
- Andrew, G., Arora, R., Bilmes, J., et al., 2013. Deep canonical correlation analysis. Int. Conf. on Machine Learning, p.1247–1255.Google Scholar
- Brownson, R.C., Gurney, J.G., Land, G.H., 1999. Evidence-based decision making in public health. J. Publ. Health Manag. Pract., 5(5)):86–97. http://dx.doi.org/10.1097/00124784-199909000-00012 CrossRefGoogle Scholar
- Carlson, C., Betteridge, J., Kisiel, B., et al., 2010. Towards an architecture for never-ending language learning. AAAI Conf. on Artificial Intelligence, p.1306–1313.Google Scholar
- Chen, D.P., Weber, S.C., Constantinou, P.S., et al., 2007. Clinical arrays of laboratory measures, or “clinarrays”, built from an electronic health record enable disease subtyping by severity. AMIA Annual Symp. Proc., p.115–119.Google Scholar
- Davenport, T.H., Prusak, L., 1998. Working Knowledge: How Organizations Manage What They Know. Harvard Business School Press, Boston, p.5.Google Scholar
- Jia, X., Gavves, E., Fernando, B., et al., 2015. Guiding long-short term memory for image caption generation. arXiv:1509.04942.Google Scholar
- Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. ImageNet: classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, p.1097–1105.Google Scholar
- Kuznetsova, P., Ordonezz, V., Berg, T.L., et al., 2014. TREETALK: composition and compression of trees for image descriptions. Trans. Assoc. Comput. Ling., 2:351–362.Google Scholar
- MIT Technology Review, 2014. Data driven healthcare. https://www.technologyreview.com/business-report/data-driven-health-care/free [Dec. 06, 2016].
- Ngiam, J., Khosla, A., Kim, M., et al., 2011. Multimodal deep learning. Int. Conf. on Machine Learning, p.689–696.Google Scholar
- Ordonez, V., Kulkarni, G., Berg, T.L., 2011. Im2text: describing images using 1 million captioned photographs. Advances in Neural Information Processing Systems, p.1143–1151.Google Scholar
- Peng, Y., Huang, X., Qi, J., 2016a. Cross-media shared representation by hierarchical learning with multiple deep networks. Int. Joint Conf. on Artificial Intelligence, p.3846–3853.Google Scholar
- Rasiwasia, N., Mahajan, D., Mahadevan, V., et al., 2014. Cluster canonical correlation analysis. Int. Conf. on Artificial Intelligence and Statistics, p.823–831.Google Scholar
- Roller, S., Schulte im Walde, S., 2013. A multimodal LDA model integrating textual, cognitive and visual modalities. Conf. on Empirical Methods in Natural Language Processing, p.1146–1157.Google Scholar
- Singhal, A., 2012. Introducing the knowledge graph: things, not strings. Official Blog of Google.Google Scholar
- Socher, R., Lin, C., Ng, A.Y., et al., 2011. Parsing natural scenes and natural language with recursive neural networks. Int. Conf. on Machine Learning, p.129–136.Google Scholar
- Socher, R., Karpathy, A., Le, Q., et al., 2014. Grounded compositional semantics for finding and describing images with sentences. Trans. Assoc. Comput. Ling., 2:207–218.Google Scholar
- Srivastava, N., Salakhutdinov, R., 2012. Multimodal learning with deep Boltzmann machines. Advances in Neural Information Processing Systems, p.2222–2230.Google Scholar
- Wu, W., Xu, J., Li, H., 2010. Learning similarity function between objects in heterogeneous spaces. Technique Report MSR-TR-2010-86, Microsoft.Google Scholar
- Xu, K., Ba, J., Kiros, R., et al., 2015. Show, attend and tell: neural image caption generation with visual attention. Int. Conf. on Machine Learning, p.2048–2057.Google Scholar
- Yang, Y., Teo, C.L., Daume, H., et al., 2011. Corpus-guided sentence generation of natural images. Conf. on Empirical Methods in Natural Language Processing, p.444–454.Google Scholar
- Zhu, Y., Zhang, C., Ré, C., et al., 2015. Building a large-scale multimodal knowledge base system for answering visual queries. arXiv:1507.05670.Google Scholar