Key-Value Memory Networks for Directly Reading Documents
Key-Value Memory Networks for Directly Reading Documents Alexander H. Miller1 Adam Fisch1 Jesse Dodge1,2 Amir-Hossein Karimi1 Antoine Bordes1 Jason Weston1 1Facebook AI Research, 770 Broadway, New York, NY, USA 2Language Technologies Institute, Carnegie Mellon University, Pittsburgh, PA, USA {ahm,afisch,jessedodge,ahkarimi,abordes,jase}@fb.com Abstract can be used to query such databases (Berant et al., 2013; Kwiatkowski et al., 2013; Fader et al., 2014). Directly reading documents and being able to Unfortunately, KBs have intrinsic limitations such answer questions from them is an unsolved as their inevitable incompleteness and fixed schemas challenge. To avoid its inherent difficulty, ques- that cannot support all varieties of answers. Since tion answering (QA) has been directed towards information extraction (IE) (Craven et al., 2000), in- using Knowledge Bases (KBs) instead, which has proven effective. Unfortunately KBs often tended to fill in missing information in KBs, is neither suffer from being too restrictive, as the schema accurate nor reliable enough, collections of raw tex- cannot support certain types of answers, and tual resources and documents such as Wikipedia will too sparse, e.g. Wikipedia contains much more always contain more information. As a result, even if information than Freebase. In this work we KBs can be satisfactory for closed-domain problems, introduce a new method, Key-Value Memory they are unlikely to scale up to answer general ques- Networks, that makes reading documents more tions on any topic. Starting from this observation, viable by utilizing different encodings in the ad- in this work we study the problem of answering by dressing and output stages of the memory read operation.
[Show full text]