Modeling multiword phrases with constrained phrase trees for improved topic modeling of conversational speech

Abstract

Latent topic modeling has proven to be an effective means for learning the underlying semantic content within document collections. Latent topic modeling has traditionally been applied to bagof-words representations that ignore word sequence information that can aid in semantic understanding. In this work we introduce a method for efficiently incorporating arbitrarily long word sequences into a topic modeling approach. This method iteratively constructs a constrained set of phrase trees in an unsupervised fashion from a document collection using weighted pointwise mutual information statistics to guide the process. In experiments on the Fisher Corpus of conversational speech, the incorporation of learned phrases into a latent topic model yielded significant improvements in the unsupervised discovery of the known topics present within the data.

Zero resource spoken audio corpus analysis

“…Its value ranges between 0 and 1, with 1 representing a perfect mapping between the true topics and the latent topics. Table 3 shows the NMI scores for a uniform random assignment of documents to latent topics, the hard agglomerative clustering used for initialization, both latent models, and a phrase-based PLSA model applied to text transcripts of the data [10]. Both models do a surprisingly good job of learning latent topics with a strong mapping to the true topics, given the fully unsupervised nature of the system.…”

Topic and Keyword Identification for Low-resourced Speech Using Cross-Language Transfer Learning

“…Make a set of all keywords for both RRL and LRL. 4. Train the Convolutional Neural Network (CNN) neural network using audio waveforms in the RRL.…”

“…Topic detection is a heavily studied problem, including methods specialized for both text [2] and speech [3] sources. Topic detection and tracking from speech is most accurately performed when one can first perform automatic speech recognition (ASR), then apply text-oriented topic detection methods such as latent Dirichlet allocation [2] or partial semantic parse [4]. It has been demonstrated that ASR-based topic detection outperforms methods without transcription, even when the ASR output has a relatively high error rate [5,6].…”