Many natural language questions are inherently personalized and subjective. They do not ask about straight facts like “who is the president of America”. Instead, they focus on personal opinions which may vary from person to person. For example, “Do you like cats?” There is no “the only correct answer” to this question. It cannot be answered properly if we do not know the personal preferences of the answerer [35].

Before this research, much effort has been devoted to the related fields. Recommender system comes straight to mind once we realize that “ratings” to different “movies” could be viewed just as the answers to different personalized questions. However, techniques used in Recommender system are usually not good at dealing with data sparsity and complex relationships. Besides, it...[ Read more ]

Many natural language questions are inherently personalized and subjective. They do not ask about straight facts like “who is the president of America”. Instead, they focus on personal opinions which may vary from person to person. For example, “Do you like cats?” There is no “the only correct answer” to this question. It cannot be answered properly if we do not know the personal preferences of the answerer [35].

Before this research, much effort has been devoted to the related fields. Recommender system comes straight to mind once we realize that “ratings” to different “movies” could be viewed just as the answers to different personalized questions. However, techniques used in Recommender system are usually not good at dealing with data sparsity and complex relationships. Besides, it is also not wise to address this problem with traditional question answering (QA) approaches. Factoid QA, using the tools of information retrieval, focuses on finding existing facts from corpus and only makes straightforward inferences. The personal traits of the answerer are not taken into consideration to give the prediction, which makes it inapplicable for this problem.

It’s hypothesized that incorporating personal traits of the answerers could help to predict their answers to personalized questions. Inspired by word representation [45], we characterize each answerer with a vector, called user representation. Both traditional and deep learning approaches have been studied. We first tried a probabilistic latent tree model built over the observations. The posterior probabilities of the internal variables of the tree are used to form the representations. However, though the traditional learning model produces meaningful and interpretable representations, the hardship brought by data sparsity remains unsettled. We, therefore, introduce a “user-question-choice” 3-modules Siamese structure to capture the preferences patterns of the answerers. The trained model is then ensembled with an additional dense layer to give predictions.