The Workshop on Feature Generation and Selection for Information Retrieval will be held on July 23, 2010, in Geneva, Switzerland, in conjunction with the 33rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2010). The workshop will bring together researchers and practitioners from academia and industry to discuss the latest developments in various aspects of feature generation and selection for textual information retrieval.
Modern information retrieval systems facilitate information access at unprecedented scale and level of sophistication. However, in many cases the underlying representation of text remains quite simple, often limited to using a weighted bag of words. Over the years, several approaches to automatic feature generation have been proposed (such as Latent Semantic Indexing, Explicit Semantic Analysis, Hashing, and Latent Dirichlet Allocation), yet their application in large scale systems still remains the exception rather than the rule. On the other hand, numerous studies in NLP and IR resort to manually crafting features, which is a laborious and expensive process. Such studies often focus on one specific problem, and consequently many features they define are task- or domain-dependent. Consequently, little knowledge transfer is possible to other problem domains. This limits our understanding of how to reliably construct informative features for new tasks.
An area of machine learning concerned with feature generation (or constructive induction) studies methods that endow computers with the ability to modify or enhance the representation language. Feature generation techniques search for new features that describe the target concepts better than the attributes supplied with the training instances. It is worthwhile to note that traditional machine learning data sets, such as those available from the UCI data repository, are only available as feature vectors, while their feature set is essentially fixed. In fact, feature generation for specific UCI benchmark datasets is scorned upon. On the other hand, textual data is almost always available in its raw format (in some case as structured data with sufficient side information). Given the importance of text as a data format, it is well worthwhile designing text-specific feature generation algorithms. Complementary to feature generation, the issue of feature selection arises. It aims to retain only the most informative features, e.g., in order to reduce noise and to avoid overfitting, and is essential when numerous features are automatically constructed. This allows us to deal with features that are correlated, redundant, or uninformative, and hence we may want to decimate them through a principled selection process.
We believe that much can be done in the quest for automatic feature generation for text processing, for example, using large-scale knowledge bases as well as the sheer amounts of textual data easily accessible today. We further believe the time is ripe to bring together researchers from many related areas (including information retrieval, machine learning, statistics, and natural language processing) to address these issues and seek cross-pollination among the different fields.
Papers from a rich set of empirical, experimental, and theoretical perspectives are invited. Topics of interest for the workshop include but are not limited to:
- Identifying cases when new features should be constructed
- Knowledge-based methods (including identification of appropriate knowledge resources)
- Efficiently utilizing human expertise (akin to active learning, assisted feature construction)
- (Bayesian) nonparametric distribution models for text (e.g. LDA, hierarchical Pitman-Yor model)
- Compression and autoencoder algorithms (e.g., information bottleneck, deep belief networks)
- Feature selection (L1 programming, message passing, dependency measures, submodularity)
- Cross-language methods for feature generation and selection
- New types of features, e.g., spatial features to support geographical IR
- Applications of feature generation in IR (e.g., constructing new features for indexing, ranking)