Fisher Kernels and Probabilistic Latent Semantic Models

Tasks that rely on semantic content of documents, notably Information Retrieval and Document Classification, can benefit from a good account of document context, i.e. the semantic association between documents. To this effect, the scheme of latent semantics blends individual words appearing throughout a document collection into latent topics, thus providing a way to handle documents that is less constrained than the conventional approach by the mere appearance of such or such word. Probabilistic latent semantic models take the matter further by providing assumptions on how the documents observed in the collection would have been generated. This allows derivation of inference algorithms that can fit the model parameters to the observed document collection; with their values set, these parameters can then be used to compute the similarities between documents. The Fisher kernels, similarity functions rooted in information geometry, constitute good candidates to measure the similarity between documents in the framework of probabilistic latent semantic models. In this context, we study the use of Fisher kernels for the Probabilistic Latent Semantic Indexing (PLSI) model. By thoroughly analysing the generative process of PLSI, we derive the proper Fisher kernel for PLSI and expose the hypotheses that relate former work to this kernel. In particular, we confirm that the Fisher information matrix (FIM) should not be approximated by the identity in the case of PLSI. We also study the impact on the performances of the Fisher kernel of the contribution of the latent topics and the one of the distribution of words among the topics; eventually, we provide empirical evidence, and theoretical arguments, showing that the Fisher kernel originally published by Hofmann, corrected to account for FIM, is the best of the PLSI Fisher kernels. It can compete with the strong BM25 baseline, and even significantly outperforms it when documents sharing few words must be matched. We further study of PLSI document similarities by applying the Language model approach. This approach shuns the usual IR paradigm that considers documents and queries to be of a similar nature. Instead, they consider documents as being representative of language models, and use probabilistic tools to determine which of these models would have generated the query with highest probability. Using this scheme in the framework of PLSI provides a way to bypass the issue of query representation, which constitutes one of the specific challenges of PLSI. We find the Language model approach to perform as well as the best of the Fisher kernels when enough latent categories are provided. Eventually, we propose a new probabilistic latent semantic model consisting in a mixture of Smoothed Dirichlet distributions which, by better modeling word burstiness, provides a more realistic model of empirical observations on real document collections than the usually used Multinomials.

Utilisation de PLSI en recherche d'information

Jean-Cédric Chappelier, Emmanuel Eckard

The PLSI model (“Probabilistic Latent Semantic Indexing”) offers a document indexing scheme based on probabilistic latent category models. It entailed applications in diverse ﬁelds, notably in information retrieval (IR). Nevertheless, PLSI cannot process documents not seen during parameter inference, a major liability for queries in IR. A method known as “folding-in” allows to circumvent this problem up to a point, but has its own weaknesses. The present paper introduces a new document-query similarity measure for PLSI based on language models that entirely avoids the problem a query projection. We compare this similarity to Fisher kernels, the state of the art similarities for PLSI. Moreover, we present an evaluation of PLSI on a particularly large training set of almost 7500 document and over one million term occurrence large, created from the TREC–AP collection.

2009

Building Word Embeddings for Solving Natural Language Processing

Rémi Philippe Lebret

Word embedding is a feature learning technique which aims at mapping words from a vocabulary into vectors of real numbers in a low-dimensional space. By leveraging large corpora of unlabeled text, such continuous space representations can be computed for capturing both syntactic and semantic information about words. Word embeddings, when used as the underlying input representation, have been shown to be a great asset for a large variety of natural language processing (NLP) tasks. Recent techniques to obtain such word embeddings are mostly based on neural network language models (NNLM). In such systems, the word vectors are randomly initialized and then trained to predict optimally the contexts in which the corresponding words tend to appear. Because words occurring in similar contexts have, in general, similar meanings, their resulting word embeddings are semantically close after training. However, such architectures might be challenging and time-consuming to train. In this thesis, we are focusing on building simple models which are fast and efficient on large-scale datasets. As a result, we propose a model based on counts for computing word embeddings. A word co-occurrence probability matrix can easily be obtained by directly counting the context words surrounding the vocabulary words in a large corpus of texts. The computation can then be drastically simplified by performing a Hellinger PCA of this matrix. Besides being simple, fast and intuitive, this method has two other advantages over NNLM. It first provides a framework to infer unseen words or phrases. Secondly, all embedding dimensions can be obtained after a single Hellinger PCA, while a new training is required for each new size with NNLM. We evaluate our word embeddings on classical word tagging tasks and show that we reach similar performance than with neural network based word embeddings. While many techniques exist for computing word embeddings, vector space models for phrases remain a challenge. Still based on the idea of proposing simple and practical tools for NLP, we introduce a novel model that jointly learns word embeddings and their summation. Sequences of words (i.e. phrases) with different sizes are thus embedded in the same semantic space by just averaging word embeddings. In contrast to previous methods which reported a posteriori some compositionality aspects by simple summation, we simultaneously train words to sum, while keeping the maximum information from the original vectors. These word and phrase embeddings are then used in two different NLP tasks: document classification and sentence generation. Using such word embeddings as inputs, we show that good performance is achieved in sentiment classification of short and long text documents with a convolutional neural network. Finding good compact representations of text documents is crucial in classification systems. Based on the summation of word embeddings, we introduce a method to represent documents in a low-dimensional semantic space. This simple operation, along with a clustering method, provides an efficient framework for adding semantic information to documents, which yields better results than classical approaches for classification. Simple models for sentence generation can also be designed by leveraging such phrase embeddings. We propose a phrase-based model for image captioning which achieves similar results than those obtained with more complex models. Not only word and phrase embeddings but also embeddings for non-textual elements can be helpful for sentence generation. We, therefore, explore to embed table elements for generating better sentences from structured data. We experiment this approach with a large-scale dataset of biographies, where biographical infoboxes were available. By parameterizing both words and fields as vectors (embeddings), we significantly outperform a classical model.

École Polytechnique Fédérale de Lausanne2016

Building Word Embeddings for Solving Natural Language Processing

Rémi Philippe Lebret

École Polytechnique Fédérale de Lausanne2016