XRCE organises public scientific seminars on a regular basis which you are welcome to attend. These seminars are an occasion to exchange with researchers from various backgrounds and to broaden scientific expertise. You can subscribe to our seminar RSS Feed for dates, speakers and topics.
Real world problems such as machine translation involve complex dependencies. Generative models have provided an elegant and flexible framework to model those dependencies, but they appear to lack robustness to model misspecification compared to discriminative models for classification. In this talk, we present methods for leveraging the advantages of generative models in the discriminative framework.
In the first part of the talk, we tackle the word alignment problem from natural language processing. We formulate it as a weighted bipartite matching problem and show how to learn the weights by using a large-margin approach for structured prediction. By providing a flexible discriminative modeling framework, we were able to cut the Alignment Error Rate in half compared to the previous best performing generative models for word alignment.
In the second part of the talk, we study probabilistic topic models which have been popular for modeling latent structures in text documents (as bag of words) or images (as bag of visual words). They are usually trained as generative models with maximum likelihood estimation, though this could be suboptimal if one is interested in doing classification. In contrast, we present a discriminative version of the Latent Dirichlet Allocation (LDA) model which attempts to uncover the latent structure in the documents while optimizing its predictive power for the task of classification. We show positive results on the 20 Newsgroup dataset for document classification.
(joint work with Fei Sha, Ben Taskar, Dan Klein and Michael I. Jordan)Status: completed
( web e-mail ), Researcher at Edelweiss project, Inria Sophia Antipolis, Sophia Antipolis, France will give a talk:
Ontology-oriented knowledge modelling and graph-based representation: a semantic web approach to manage communities and their knowledge
Using motivating scenarios we will first position this talk in the domain of knowledge representation focusing on ontology-based collective memories. We will show how formalisms based on graphs can be used to represent knowledge with varying degrees of formalization depending on the needs identified in the application scenarios and processing they require. We will focus on the graph-oriented models of the semantic web. We will also defend the idea that ontologies support other types of reasoning than logical derivation and in particular that the graphs of these knowledge representations can be seen as metric spaces used to pilot approximation, or index of knowledge in distributed environments, or sources to make interfaces more intelligible to end-users, or new frameworks for social network analysis.Slides (8.79 MB)
Significant improvements have been achieved in machine translation over the past few years. It is mostly motivated by appearance of statistical machine translation (SMT) technology that is currently considered as the best way to do automatic translation of natural languages.
This talk focuses on a syntax-based approach to handle the fundamental problem of word ordering for SMT exploiting syntactic representations of source and target texts. The talk begins with the existing idea of taking reordering rules automatically derived through a syntactically augmented alignment of source and target texts. A new approach to hierarchically extract reordering patterns is then proposed. A set of extracted reordering rules is applied in a preprocessing step before translation to make the source sentence structurally more like the target. We evaluate the proposed approach, combined with a POS lattice-based decoding, on the Arabic-to-English and Chinese-to-English translation tasks.
Furthermore, a brief introduction to an N-gram-based approach to SMT will be given, along with analysis of major differences between N-gram-based and phrase-based approaches to SMT.
The talk will be easily accessible to a wide audience.Slides (1.51 MB)
In this talk I will present our work on British Sign Language (BSL) recognition. Specifically, I will (i) show how we detect the pose of a signer (arms, head and body) to find the position of the hands and (ii) demonstrate that BSL signs can be learned automatically using signing footage and simultaneously broadcasted subtitles taken from TV.
Detecting the pose of a signer is cast as inference in a generative model of the image. Under this model, limb detection is expensive due to the very large number of possible configurations each part can assume. We make the following contributions to reduce this cost: (i) using efficient sampling from a pictorial structure proposal distribution to obtain reasonable configurations; (ii) identifying a large set of frames where correct configurations can be inferred, and using temporal tracking elsewhere.
Learning BSL signs automatically from TV broadcasts is achieved using the supervisory information available from subtitles broadcast simultaneously with the signing. We propose (i) a distance function to match signing sequences which includes the trajectory of both hands, the hand shape and orientation; (ii) we show that by optimizing a scoring function based on multiple instance learning, we are able to extract the sign of interest from hours of signing footage, despite the very weak and noisy supervision.Slides (6.90 MB)
Global competitive pressures and rapid technological change make organizations a dynamic environment where workers are forced to continuously update their work-related skills and practices. Work and learning are becoming increasingly equivalent (e.g., Tapscott 1995). This convergence is particularly evident in complex, non-routine knowledge work such as the kind involved in crafting and selling services, which is part of our focus. Successfully selling, implementing, and delivering 'customized' services to clients requires highly skilled work; experience matters. Moreover, this type of industry appears more exposed to risk from losses of intellectual capital if companies fail to address the problem of knowledge handoff between retiring seniors and junior colleagues (e.g., DeLong 2004). As trends for the corporate IT industry show that the business of providing online services has steadily increased since the 1990s (Cusumano 2008), the problem of supporting professional development and knowledge exchange in this context is likely to become more evident in the future. Forming communities of practices for supporting learning and development in the workplace represents a promising solution to this problem. In fact, communities of practice are being promoted within many organizations as a way to build competences and social capital: intra-organization groups of workers that share interests, problems, and tips (e.g., Wenger 2002, Probst & Borzillo 2008). We study those explicitly aimed at professional development: i.e. learning communities. While Web 2.0 tools are being adopted by global enterprises, we study computer-supported communities of practice where services professionals with similar job functions in a large enterprise can share knowledge and develop skills. Although a long tradition of field studies has focused on the practices and tools for exchanging expertise within organizations, the literature is lacking systematic investigations of how learning communities form and what causes their success or failure (see the classic study of Xerox's field service technicians sharing tips reported by Orr 1996). A grand challenge for social computing researchers is to first develop understanding and then translate it into tools that successfully foster the growth of learning communities in a corporate setting (see how the findings about technicians' practices led Xerox to implement the Eureka system, or the current findings and lessons learned about building online communities (Kim 2000, Iriberri and Leroy 2009).
A number of qualitative observations from our intra-organization communities, consistently with prior success cases of online communities, suggest that the design of community-building interventions (i.e., web tools and development programs) are likely to be sustainable when led by an indirect design approach. That is, when the tools and the programs help directly with the daily work of individuals and their task forces (rather than being an additional burden for them) and at the same time enable as a side effect the new opportunity for individuals to easily contribute to community resources while doing the daily work (e.g., wiki pages, shared bookmarks, shared annotations). Inspired by this approach, our strategy to design better support for our learning communities starts from the identification of key needs of the individual professionals and their task forces. We believe that by adequately addressing these we can create specific conditions that promote the emergence of communities.
Overall, the members of the two communities investigated had consistent views about some basic needs that technology can address to better support their work. Two key problems that the services professionals face regularly are handling (a) a large amount of information which they need to monitor or process daily and (b) many information channels which they need to coordinate (email, phone, DocuShare, wikis). These channels currently act independently from one another resulting in an extra cost for the users. Support for learning and development, such as the technology introduced, should help them address, at least in part, these two challenges.Slides (3.40 MB)