Learning to refine text based recommendations

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2103–2108, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Learning to refine text based recommendations Youyang Gu and Tao Lei and Regina Barzilay and Tommi Jaakkola Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology {yygu,taolei,regina,tommi}@csail.mit.edu Abstract We propose a text-based recommendation en- gine that utilizes recurrent neural networks to flexibly map textual input into continuous vec- tor representations tailored to the recommen- dation task. Here, the text objects are doc- uments such as Wikipedia articles or ques- tion and answer pairs. As neural models re- quire substantial training time, we introduce a sequential component so as to quickly ad- just the learned metric over objects as addi- tional evidence accrues. We evaluate the ap- proach on recommending Wikipedia descrip- tions of ingredients to their associated product categories. We also exemplify the sequential metric adjustment on retrieving similar Stack Exchange AskUbuntu questions. 1 1 Introduction Modern recommender problems involve complex objects, often described in textual form. In order to learn to predict how disparate objects may go to- gether, it is helpful to first map them into a common representation where they are easily compared, re- gardless of their origin. Neural models are partic- ularly well-suited for this task as continuous vec- tor representations of objects can be tailored in a flexible way to the desired task. While these mod- els have been shown to be effective across NLP tasks (Sutskever et al., 2014; Andreas et al., 2016; Hermann et al., 2015), they take considerable time to learn and are therefore ill-suited to be adjusted rapidly as additional evidence accumulates. 1The code/data is available at https://github.com/ youyanggu/rcnn. We cast our text-to-text recommendation problem in two phases. In the first phase, flexible neural text- to-vector mappings are learned from currently avail- able data. Such mappings are optimized to function well in a collaborative filtering setting. For exam- ple, in the context of recommending food product categories for ingredients based on their Wikipedia pages, the continuous vectors are adjusted so that their inner product directly reflects the degree of as- sociation between the objects. Once learned, the mapping can be applied to any previously unseen text to yield the corresponding vector representation, and therefore also used for predicting associations. In the second phase, we no longer adjust text-to- vector mappings but rather parameterize and learn how the vectors are compared. For example, we can optimize the metric separately for each new ingredi- ent based on a few category observations for that in- gredient. The goal of this second phase is to specif- ically boost the accuracy when the neural baseline (unaware of the new evidence) would otherwise not perform well. Our approach builds on the recent work on recur- rent convolutional models to obtain text-to-vector mappings (Lei et al., 2015; Lei et al., 2016). This architecture is particularly well suited for noisy Wikipedia pages as it can learn to omit and high- light different parts of …

Content truncated to 500 words