User intent prediction in information-seeking on workflow prediction is defined as follows. The input of the system is an information-seeking dialog
dataset D = {(Ui , Yi)}N i=1 and a set of user intent labels L = {l1,l2, . . . ,lc }. A dialog Ui = {u 1 i ,u 2 i , . . . ,u k i }
contains multiple turns of utterances. u k i is the utterance at the k-th turn of the i-th dialog. Yi consists of annotated user intent labels
{y 1 i , y 2 i , . . . , y k i }, where y k i = {y k(1) i ,y k(2) i , . . . ,y k(c) i }. Here y k(m) i , . . . ,y k(n) i = 1 denotes that the utterance u k i in dialog Ui is labeled with user intent {lm, . . . ,ln}.
Given an utterance u k i and other utterances in dialog Ui , the goal is to predict the user intent Yi of this utterance.
The challenge of this task lies in the complexity and diversity of human information-seeking conversations, where one utterance often expresses multiple user intent separation.
To apply traditional ML methods with features to this task, we need to transform this multi-label classification problem to multi-class classification.
Multiple transformation strategies are typically used: binary relevance, classifier chains, and label powerset - Binary relevance does not consider the label correlations and label powerset generates new labels for every label combination.
So we choose classifier chains as the transformation strategy for traditional ML methods. This strategy performs binary classifications for each label and take predictions for previous labels as extra features.
This transformation strategy is the best fit for our task as it considers the label dependency without explicitly generating new labels for every label combination. We adopt classic ML methods, including Naive Bayes classifier, SVM, random forest, and AdaBoost as baseline classifiers. In addition, we use ML-kNN, which supports multi-label classification by nature.