Translation of Natural Language Query Into Keyword Query Using a RNN Encoder-Decoder

A Reinforcement Learning-driven Translation Model for Search-Oriented Conversational Systems

“…It is worth mentioning that we also trained in preliminary experiments a state of the art translation models such as a generative encoder-decoder RNN with attention mechanism, as done in (Yin et al, 2017;Song et al, 2017). We did not report the results since the model was not able to generalize in the testing phase over new samples from the NL-query dataset used in the training phase.…”

“…We are aware that the use of TREC datasets is biased in the sense that it does not exactly fit with the expression of NL information need in the context of conversational systems, but we believe that the description is enough verbose to evaluate the impact of our query building model in this ex- (Song et al, 2017), will be carried out in the future. We also analyze the issue of duplicate words into TREC descriptions since it can directly impact the query formulation process based on word selection in the word sequence of TREC descriptions.…”

“…• SMT which only considers the first component of our model based on a supervised machine translation approach (Section 2.2). This variant could be assimilated to the approach proposed in (Song et al, 2017) in the sense that the machine translation is performed independently of the task objective.…”

“…In SOCS, there is an upstream challenge dealing with the building of the query from a NL expression that initiates the search session to avoid useless users' interactions with the system. This problem could be tackled for instance through deep neural translation models (e.g., encoder-decoder approaches) as initiated by (Song et al, 2017;Yin et al, 2017). However, these methods learn the query formulation model independently of the search task at hand.…”

Reformulating natural language queries using sequence-to-sequence models

“…This "lexical chasm" problem has attracted increasing attention from both academic and enterprise communities. Query reformulation, which attempts to alleviate vocabulary mismatch by changing a given natural language query into an alternative query, is considered as an effective solution for this problem, and various related approaches have been proposed [1][2][3][4]. For example, Jones et al [2] proposed generating a new query to replace the natural language query and considered query reformulation as a machine translation problem, where user queries are treated as one language and the corresponding reformulated queries are treated as another language.…”