Model Training

During the training phase we build a model of conversational topics which is used for classification. The model in our system is represented by a multinomial probability distribution computed from the counts of unique words for each topic:

$$P(word_i\vert c) = \frac{N_{c}(word_i)}{\sum_{j = 1}^{w}N_{c}(word_j)}$$ (1)

where P_c(word_j | c) is a probability of a particular word in the class c, N_c(word_j) is how many times this word was encountered in the training corpus for the class c and w is the total number of unique words in the class. This model is a word frequency model which converges to true word probabilities given large training corpus.

Currently we train the system for 12 different conversation topics. The training data is a corpus of text documents collected from the web and newsgroups. The current topics include the following:

  Newsgroup: alt.business - business related discussions
  Newsgroup: alt.jobs - job announcements
  Newsgroup: alt.politics.british - discussions of British politics
  Newsgroup: alt.politics.clinton - discussions about President Clinton
  Newsgroup: alt.politics.economics - economic discussions
  Newsgroup: comp.ai.philosophy - discussions on Artificial Intelligence
  WWW site:  deptofstate - documents from US Department of State
  WWW site:  intlcourtofjustice - docs from International Court of Justice
  Newsgroup: pilot.programmer - Programming Pilot PDA
  Newsgroup: soc.singles - dating and relationships
  Newsgroup: talk.politics.medicine - discussions about medicine
  Newsgroup: talk.religion.misc - discussions about religion

No lexical filtering is performed to model word distributions within each topic. Ideally, we need to have spoken conversational data collected in the same setting as the expected interaction to increase accuracy. However, text sources provide a good starting point for testing our algorithm.