Generating training datasets for a supervised learning topic model from outputs of a discovery topic model
Systems and methods for generating training data for a supervised topic modeling system from outputs of a topic discovery model are described herein. In an embodiment, a system receives a plurality of digitally stored call transcripts and, using a topic model, generates an output which identifies a plurality of topics represented in the plurality of digitally stored call transcripts. Using the output of the topic model, the system generates an input dataset for a supervised learning model by identify a first subset of the plurality of digitally stored call transcripts that include the particular topic, storing a positive value for the first subset, identifying a second subset that do not include the particular topic, and storing a negative value for the second subset. The input training dataset is then used to train a supervised learning model.
1. A computer-implemented method comprising:
receiving a plurality of digitally stored call transcripts that have been prepared from digitally recorded voice calls;
based on the call transcripts, creating and storing a topic model, the topic model simultaneously modeling topics for words in a call as a first function of one or more topic distributions for each topic of a plurality of topics that are used over a plurality of calls, and words of the call as a second function of call-specific probability distribution of words for each topic of the plurality of topics used over the plurality of calls;
the one or more topic distributions modeling each of the one or more topic distributions as a third function of an inferred prior probability distribution that is modeled as a fourth function of a flat prior distribution;
each call-specific probability distribution of words being modeled as a fifth function of a probability distribution of words that is modeled as a sixth function of an inferred prior distribution modeled as a seventh function of a flat prior distribution;
based on output of the topic model, generating an input dataset for a supervised learning model by:
identifying a first subset of the plurality of digitally stored call transcripts that include a particular topic, the first subset of the plurality of digitally stored call transcripts being stored with a positive out value;
identifying a second subset of the plurality of digitally stored call transcripts that do not include the particular topic, the second subset of the plurality of digitally stored call transcripts being stored with a negative value;
determining that a second topic of the plurality of topics corresponds to the particular topic by:
determining that a particular probability distribution of words for the particular topic is greater than a second probability distribution for the second topic multiplied by a value corresponding to a length of the call;
identifying one or more digitally stored call transcripts of the second subset of the plurality of digitally stored call transcripts that include the second topic;
removing the one or more digitally stored call transcripts from the input dataset; and
training the supervised learning model using the generated input dataset, wherein the supervised learning model is configured to compute, for a new digitally stored call transcript, a probability that the particular topic was discussed during a digitally recorded voice call corresponding to the new digitally stored call transcript.
2. The computer-implemented method of claim 1 , further comprising, in determining that the particular probability distribution of words for the particular topic is greater than the second probability distribution for the second topic multiplied by a value corresponding to a length of the call, the value corresponding to the length of the call being calculated as
θ
dk
≥
α
k
(
N
d
)
2
3
where is an average call length for the plurality of calls and θ dk represents the second probability distribution.
3. The computer-implemented method of claim 2 , further comprising determining whether the call transcript includes the particular topic by:
storing a word number threshold for a number of words in the call transcript that correspond to the particular topic;
determining that the number of words in the call transcript that correspond to the particular topic is greater than the word number threshold, and in response, determining that the call transcript includes the particular topic.
4. The computer-implemented method of claim 1 , further comprising determining whether the call transcript includes the particular topic by:
storing a word number threshold for a number of words in the call transcript that correspond to the particular topic;
determining that the number of words in the call transcript that correspond to the particular topic is greater than the word number threshold, and in response, determining that the call transcript includes the particular topic.
5. The computer-implemented method of claim 1 , further comprising determining that words in the call transcript correspond to the particular topic by drawing a sample from the second probability distribution or an augmented prior probability distribution.
6. The computer-implemented method of claim 5 , further comprising:
calculating an expected number of words on each call d associated with topic k as ρ dk =N d *θ dk ;
drawing a sample from a distribution for the expected number of words on the call (ρ dk );
identifying each unique word in the call transcript that matches the sample (z kd );
identifying the call transcript as including the particular topic only if the number of unique words in the call that matches the sample (z kd ) exceeds a word number threshold.
7. The computer-implemented method of claim 6 , further comprising drawing a same sample for each call transcript.
8. A computer system comprising:
one or more hardware processors;
one or more non-transitory storage media coupled to the one or more hardware processors and storing one or more sequences of instructions which, when executed using the one or more hardware processors, cause the one or more hardware processors to execute:
receiving a plurality of digitally stored call transcripts that have been prepared from digitally recorded voice calls;
based on the call transcripts, creating and storing a topic model, the topic model simultaneously modeling topics for words in a call as a first function of one or more topic distributions for each topic of a plurality of topics that are used over a plurality of calls, and words of the call as a second function of call-specific probability distribution of words for each topic of the plurality of topics used over the plurality of calls;
the one or more topic distributions modeling each of the one or more topic distributions as a third function of an inferred prior probability distribution that is modeled as a fourth function of a flat prior distribution;
each call-specific probability distribution of words being modeled as a fifth function of a probability distribution of words that is modeled as a sixth function of an inferred prior distribution modeled as a seventh function of a flat prior distribution;
based on output of the topic model, generating an input dataset for a supervised learning model by:
identifying a first subset of the plurality of digitally stored call transcripts that include a particular topic, the first subset of the plurality of digitally stored call transcripts being stored with a positive out value;
identifying a second subset of the plurality of digitally stored call transcripts that do not include the particular topic, the second subset of the plurality of digitally stored call transcripts being stored with a negative value;
determining that a second topic of the plurality of topics corresponds to the particular topic by:
determining that a particular probability distribution of words for the particular topic is greater than a second probability distribution for the second topic multiplied by a value corresponding to a length of the call;
identifying one or more digitally stored call transcripts of the second subset of the plurality of digitally stored call transcripts that include the second topic;
removing the one or more digitally stored call transcripts from the input dataset; and
training the supervised learning model using the generated input dataset, wherein the supervised learning model is configured to compute, for a new digitally stored call transcript, a probability that the particular topic was discussed during a digitally recorded voice call corresponding to the new digitally stored call transcript.
9. The computer system of claim 8 , further comprising sequences of instructions which, when executed using the one or more hardware processors, cause the one or more hardware processors to execute, in determining that the particular probability distribution of words for the particular topic is greater than the second probability distribution for the second topic multiplied by a value corresponding to a length of the call, the value corresponding to the length of the call being calculated as
θ
dk
≥
α
k
(
N
d
)
2
3
where is an average call length for the plurality of calls and θ dk represents the second probability distribution.
10. The computer system of claim 9 , further comprising sequences of instructions which, when executed using the one or more hardware processors, cause the one or more hardware processors to execute determining whether the call transcript includes the particular topic by:
storing a word number threshold for a number of words in the call transcript that correspond to the particular topic;
determining that the number of words in the call transcript that correspond to the particular topic is greater than the word number threshold, and in response, determining that the call transcript includes the particular topic.
11. The computer system of claim 8 , further comprising sequences of instructions which, when executed using the one or more hardware processors, cause the one or more hardware processors to execute determining whether the call transcript includes the particular topic by:
storing a word number threshold for a number of words in the call transcript that correspond to the particular topic;
determining that the number of words in the call transcript that correspond to the particular topic is greater than the word number threshold, and in response, determining that the call transcript includes the particular topic.
12. The computer system of claim 8 , further comprising sequences of instructions which, when executed using the one or more hardware processors, cause the one or more hardware processors to execute determining that words in the call transcript correspond to the particular topic by drawing a sample from the second probability distribution or an augmented prior probability distribution.
13. The computer system of claim 12 , further comprising sequences of instructions which, when executed using the one or more hardware processors, cause the one or more hardware processors to execute:
calculating an expected number of words on each call d associated with topic k as ρ dk =N d *θ dk ;
drawing a sample from a distribution for the expected number of words on the call (ρ dk );
identifying each unique word in the call transcript that matches the sample (z kd );
identifying the call transcript as including the particular topic only if the number of unique words in the call that matches the sample (z kd ) exceeds a word number threshold.
14. The computer system of claim 13 , further comprising sequences of instructions which, when executed using the one or more hardware processors, cause the one or more hardware processors to execute drawing a same sample for each call transcript.