Latent Topic-based Subspace for Natural Language Processing

Natural Language Processing (NLP) applications have difficulties in dealing with automatically transcribed spoken documents recorded in noisy conditions, due to high Word Error Rates (WER), or in dealing with textual documents from the Internet, such as forums or micro-blogs, due to misspelled or truncated words, bad grammatical form… To improve the robustness against document errors, hitherto-proposed methods map these noisy documents in a latent space such as Latent Dirichlet Allocation (LDA), supervised LDA and author-topic (AT) models. In comparison to LDA, the AT model considers not only the document content (words), but also the class related to the document. In addition to these high-level representation models, an original compact representation, called c -vector, has recently been introduced avoid the tricky choice of the number of latent topics in these topic-based representations. The main drawback in the c -vector space building process is the number of sub-tasks required. Recently, we proposed both improving the performance of this c -vector compact representation of spoken documents and reducing the number of needed sub-tasks, using an original framework in a robust low dimensional space of features from a set of AT models called “Latent Topic-based Subspace” (LTS). This paper goes further by comparing the original LTS-based representation with the c -vector technique as well as with the state-of-the-art compression approach based on neural networks Encoder-Decoder (Autoencoder) and classification methods called deep neural networks (DNN) and long short-term memory (LSTM), on two classification tasks using noisy documents taking the form of speech conversations but also with textual documents from the 20-Newsgroups corpus. Results show that the original LTS representation outperforms the best previous compact representations with a substantial gain of more than 2.1 and 3.3 points in terms of correctly labeled documents compared to c -vector and Autoencoder neural networks respectively. An optimization algorithm of the scoring model parameters is then proposed to improve both the robustness and the performance of the proposed LTS-based approach. Finally, an automatic clustering approach based on the radial proximity between documents classes is introduced and shows promising performances.