Siks Course on Learning and Information Retrieval (may 2007) Michiel Hildebrand First some ideas that got shape during the course. They are mainly about adding probabilities to the process of constructing graph paths. The idea of probability reasoning given a model is applicable to many situations. In speech recognition and IR a language model is used to predict the probability of a word given a chain. I like to further study how we could use this approach to compute probabilities of the next property-value pair in a chain of relations. For example, given the chain "Paris -birthPlace- Person" what is the probability of an extension to "Paris -birthPlace- Person -creator- Artwork". The problem of course is the model that we need for this. I think the requirement is a "model of use". We claim to have semantics in our data. I am still not convinced, for me, semantics is related to action. To qoute Wittgenstein "meaning is use". The formal semantics of RDFS and OWL describe how to use the constructs of these language, however, this only applies in a particular setting. A setting in which we assume the (universal) model/ontology applies to all users and formal queries are used to access information. In a setting with many different types of users that have vague information needs we have to be much more pragmatic. What the semantics are depends on the user. So, in e-Culture, and many other practical SW applications, we have linked data and we still need to determine and capture the semantics. In other words, we need to capture how to use predicates and concepts given a particular user need. This is our model of use. I think/hope that in this model there is room for formal logic as well as probability theory. Instead of individual results we can move towards paths connecting various related results. Users can construct such paths by searching and browsing. In some literature this is known as a trail. In a web3.0 setting users could construct such trails deliberately, annotate them with a name, description, comments etc. We could also monitor them behind the screen. Qualitative Reasoning (Bert Bredeweg) ------------------------------------- Qualitative aspects are important for communication of decision procedure. We need a vocabulary to communicate. Modeling tries to capture knowledge chunks -> the essence in a certain domain Collaboratively we use the Garp3, a tool for modeling and reasoning over situations (developed for ecology). We model the connected vessels problem using a type hierarchy, the available properties with a qualitative value range (0,pos,max), a configuration with instances of the types, the properties of the instances and all dependencies between instances and properties. The system then generates a behavior graph from the possible qualitative state given an initial state. Nice example, because we modeled the possible value of the amount to be max, there are two unexpected states, namely, It flows over or levels exactly at max. Metaheuristic search for combinatorial optimization (Dirk Thierens) ------------------------------------------------------------------- Local search is an efficient set of algorithms for combinatorial problems, such as the knapsack problem. Start with a set of results (possibly random) and use a neighborhood function to pick new results and replace initial result if it is better. Multi start local search performs multiple random local searches. Meta-heuristic search tries to optimize the random picking procedure. Uncertainty reasoning with bayesian networks (Peter Lucas) ---------------------------------------------------------- This was interesting and I will further study this. Use probability theory but allows reasoning over the structure of the graph. 1 local probabilities 2 independence representation in graphs There are computational optimizations, for example, Judea Pearl (message passing through communication channels which are the arcs). Traditionally baysian networks are propositional of nature. There is work done to model predicate logic as well, for example, bayesian (probabilistic) logic and relational learning. Automated reasoning for AI (Stefan Schlobach) --------------------------------------------- Nice enthusiastic guy! Automated reasoning is logic in action. It bridges the theory and the real world. He explains basic Automated Reasoning and the Davis Putman procedure. The rest of the talk is an intro into Description Logics. Formal models of IR (Djoerd Hiemstra) ------------------------------------- Introduction to 5 models for document retrieval using full text. 1 boolean retrieval only models matching terms 2 vector space model (tf.idf) - the query as a document. best documents are similar to queries (luhn, 1957) - vector is used instead of a point because it normalizes how far away a document is through changes in document size. 3 probability ranking - requires set of relevant document 4 language models random sampling words with background language knowledge 5 pagerank random surfer Language models (Wessel Kraai) ------------------------------ Originated in speech recognition. Probability of a term given a memory (chain rule) P(fellow|"for he is a jolly") = word order modeling. The rest is statistical blabla. Multimedia IR (Arjen de Vries) ------------------------------ Feedback is crucial for content based retrieval. Available index data: - manually assigned metadata - associated data, speech, captions - content-based He describes problems with using only metadata: the vocabulary used, encoding specificity, some things can't be expressed (unambiguously) in language. 'Query by example' -- similarity based -- vector space model with features and compare. Works for some examples, works terrible for many others. Arnold Smeulders solution. Domain specific detectors. High level concepts which can be recognized at high accuracy with special selectors. Arjen's phd work: user feedback on local information and global information (click graph). Create clusters in vector space. Clusters correspond to concepts. The idea of a click graph is very interesting and I am playing with this thoughts for months. How can we can information about interesting graph paths by monitoring click in a user session. Language models for MM. In generative mode, we ask what is the model that could generate the image. -> capturing the essence of an image in a gaussian mixture model. Question answering (David Ahn) ------------------------------ Nice talk about question answering. No comments about the talk. During the break David told me he is going back to the States and work for a company called Powerset. (http://www.powerset.com/). They plan to parse the whole web, extract relations and built a search engine on top of that. Information Retrieval --------------------- I took two things out of this. With an IR model we can try to improve how the user's information need is expressed with a query. a-prior: user learns syntax & semantics of query language (with the formal query as a far end) a-posteriori: interaction manager (interactively built query) There is a distinction between extensional representations and intensional. The index of a document is extensional, whereas, the user's query representing the users information need is intensional. It's easier to judge relevance of a result (extensional) than how good the query represents the information need (intensional). Therefore, we should allow the user to give feedback on the results.