Unsupervised topic modeling for short texts
Topics are determined for short text messages using an unsupervised topic model. In a training corpus created from a number of short text messages, a vocabulary of words is identified, and for each word a distributed vector representation is obtained by processing windows of the corpus having a fixed length. The corpus is modeled as a Gaussian mixture model in which Gaussian components represent topics. To determine a topic of a sample short text message, a posterior distribution over the corpus topics is obtained using the Gaussian mixture model.
1. A method, comprising:
by a computer, obtaining distributed vector representations of words in a vocabulary identified in a corpus comprising a plurality of training short text messages, the distributed vector representations being obtained by processing context windows of the corpus using a continuous bag of words model, the continuous bag of words model individually processing each one of the context windows and obtaining an individual distributed vector representation for each one of the context windows;
by the computer, estimating a plurality of Gaussian components of a Gaussian mixture model of the corpus using the distributed vector representations, the Gaussian components representing corpus topics, wherein the estimating the plurality of Gaussian components of the Gaussian mixture model of the corpus further comprises using bottleneck features obtained using neural networks;
by the computer, receiving a sample short text message comprising a subset of the words in the vocabulary; and
by the computer, determining a topic of the sample short text message based on a posterior distribution over the corpus topics for the sample short text message, the posterior distribution obtained using the Gaussian mixture model.
2. The method of claim 1 , wherein the training short text messages have a maximum message length, and the context windows have a fixed length greater than or equal to the maximum message length.
3. The method of claim 1 , wherein the using the continuous bag of words model further comprises using a log-linear model.
4. The method of claim 1 , wherein the obtaining the distributed vector representations of the words in the vocabulary further comprises applying a methodology to process the context windows of the corpus, the methodology being selected from a group of methodologies consisting of deep neural network, latent semantic indexing, log-linear model, feedforward neural network, convolutional neural network and recurrent neural network.
5. The method of claim 1 , further comprising:
reducing the vocabulary using hierarchical sampling.
6. The method of claim 5 wherein the hierarchical sampling eliminates words having fewer than five occurrences.
7. The method of claim 1 wherein the training short text messages have an average text length of between 12 and 16 words.
8. The method of claim 1 wherein the estimating the plurality of Gaussian components further comprises estimating means, covariances and mixture weights for each Gaussian component using an expectation-maximization algorithm.
9. The method of claim 8 wherein the covariances are estimated using a covariance matrix approximation wherein the covariances are diagonal matrices.
10. The method of claim 1 wherein the posterior distribution over the corpus topics for the sample short text message is determined by evaluating:
k
*
=
arg
max
θ
p
p
(
k
)
∏
i
=
1
N
p
(
w
i
′
❘
k
)
where k′ is a posterior distribution for a topic k, θ k denotes parameters for the k th Gaussian component of the Gaussian mixture model, w′ i is the i th word in the sample short text message and the probabilities p(k) and p(w′ i |k) are obtained from the Gaussian mixture model.
11. The method of claim 1 , wherein the vocabulary in the corpus is identified by representing a phrase of words within the corpus by a single code word to minimize a description length of the corpus.
12. A system of a communications network, comprising:
an interface to the communications network configured for receiving short text messages transmitted within the communications network;
a processor; and
a computer readable storage device having stored thereon computer readable instructions that, when executed by the processor, cause the processor to perform operations, the operations comprising:
obtaining distributed vector representations of words in a vocabulary identified in a corpus comprising a plurality of training short text messages, the distributed vector representations being obtained by processing context windows of the corpus using a continuous bag of words model, the continuous bag of words model individually processing each one of the context windows and obtaining an individual distributed vector representation for each one of the context windows;
estimating a plurality of Gaussian components of a Gaussian mixture model of the corpus using the distributed vector representations, the Gaussian components representing corpus topics, wherein the estimating the plurality of Gaussian components of the Gaussian mixture model of the corpus further comprises using bottleneck features obtained using neural networks;
receiving a plurality of sample short text messages comprising a subset of the words in the vocabulary;
determining topics of the sample short text messages based on a posterior distribution over the corpus topics for the sample short text messages, the posterior distribution obtained using the Gaussian mixture model;
identifying a trend in the topics of the sample short text messages; and
generating an alert based on the trend.
13. The system of claim 12 , wherein the sample short text messages have a maximum message length, and the context windows have a fixed length greater than or equal to the maximum message length.
14. The system of claim 12 , wherein the sample short text messages have an average message length, and the context windows have a fixed length that approximates the average message length.
15. The system of claim 12 , wherein the using the continuous bag of words model further comprises using a log-linear model.
16. The system of claim 12 wherein the operations further comprise:
reducing the vocabulary using hierarchical sampling.
17. The system of claim 12 wherein the estimating the plurality of Gaussian components further comprises estimating means, covariances and mixture weights for each Gaussian component using an expectation-maximization algorithm.
18. The system of claim 12 , wherein the vocabulary identified in the corpus is identified by representing a phrase of words within the corpus by a single code word to minimize a description length of the corpus.
19. A tangible computer-readable medium having stored thereon computer readable instructions, wherein execution of the computer readable instructions by a processor causes the processor to perform operations, the operations comprising:
obtaining distributed vector representations of words in a vocabulary identified in a corpus comprising a plurality of training short text messages, the distributed vector representations being obtained by processing context windows of the corpus using a continuous bag of words model, the continuous bag of words model individually processing each one of the context windows and obtaining an individual distributed vector representation for each one of the context windows;
estimating a plurality of Gaussian components of a Gaussian mixture model of the corpus using the distributed vector representations, the Gaussian components representing corpus topics, wherein the estimating the plurality of Gaussian components of the Gaussian mixture model of the corpus further comprises using bottleneck features obtained using neural networks;
receiving a sample short text message comprising a subset of the words in the vocabulary; and
determining a topic of the sample short text message based on a posterior distribution over the corpus topics for the sample short text message, the posterior distribution obtained using the Gaussian mixture model.
20. The tangible computer-readable medium of claim 19 , wherein the training short text messages have a maximum message length, and the context windows have a fixed length greater than or equal to the maximum message length.