Section: New Results

Unsupervised motif discovery

Knowledge discovery aims at bringing out coherent groups of objects and is usually based on clustering which necessitates defining a notion of similarity between objects. In practice, this strong prior is often neither possible nor welcome. We proposed to divert supervised machine learning (ML) techniques in order to calculate, indirectly and without supervision, similarities among objects. Our approach consists in generating artificial labeling problems on the data to reveal regularities between objects through their labeling. In [28] , we show how this framework can be implemented and experimented on two information extraction/discovery tasks concerned with named entities. The ML technique diverted to exhibit similarities between with the named entities are the Conditional Random Fields. This same method can also be applied with less common ML techniques: In [59] , we show that Inductive Logic Programming can also be used to cluster complex data. Thanks to the ability of ILP to handle data that cannot be expressed under the usual attribute-value representation, we use it to make emerge clusters of TV broadcasts based only on their broadcasting information (date, time, length, etc.).

We evaluated a system to create audio thumbnails of spoken content, i.e., short audio summaries representative of the entire content, without resorting to a lexical representation. As an alternative to searching for relevant words and phrases in a transcript, unsupervised motif discovery is used to find short, word-like, repeating fragments at the signal level without acoustic models. The output of the word discovery algorithm is exploited via a maximum motif coverage criterion to generate a thumbnail in an extractive manner. A limited number of relevant segments are chosen within the data so as to include the maximum number of motifs while remaining short enough and intelligible. Evaluation is performed on broadcast news reports with a panel of human listeners judging the quality of the thumbnails. Results indicate that motif-based thumbnails stand between random thumbnails and ASR-based keywords, however still far behind thumbnails and keywords humanly authored [34] .

Unsupervised approaches were introduced a few years ago to analyze the structure of TV programs, relying on the discovery of repeated elements within a program or across multiple episodes of the same program. These methods can discover key repeating elements, such as jingles and separators, however they cannot infer the entire structure of a program. In [48] , we studied a hierarchical use of grammatical inference to yield a temporal grammar of a program from a collection of episodes, discovering both the vocabulary of the grammar and the temporal organization of the words from the vocabulary. Using a set of basic event detectors and simple filtering techniques to detect repeating elements of interest, a symbolic representation of each episode is derived based on minimal domain knowledge. Grammatical inference based on multiple sequence alignment is then used in a hierarchical manner to provide a temporal grammar of the program at various levels of details.

Many nearest neighbor search algorithms rely on encoding real vectors into binary vectors. The most common strategy projects the vectors onto random directions and takes the sign to produce so-called sketches. In [22] , we discuss the sub-optimality of this choice, and propose a better encoding strategy based on the quantization and reconstruction points of view. Our second contribution is a novel asymmetric estimator for the cosine similarity. Similar to previous asymmetric schemes, the query is not quantized and the similarity is computed in the compressed domain. We tackled the same similarity estimation problem with a rather different approach in [52] , where we assume that only a few vectors of the database, so-called heavy hitters, have a similarity to the query that significantly deviates from 0. For this purpose, we have introduced a group testing framework for detecting large similarities between high-dimensional vectors, such as descriptors used in state-of-the-art description of multimedia documents. We produce a set of group representations that jointly encode several vectors into a single one, in the spirit of group testing approaches. By comparing a query vector to several of these intermediate representations, we screen the large values taken by the similarities between the query and all the vectors, at a fraction of the cost of exhaustive similarity calculation. Unlike concurrent indexing methods that suffer from the curse of dimensionality, our method exploits the properties of high-dimensional spaces.