Hi again all, I want to compare translation results with my tuned Japanese-English system, and the same system with the lexical reordering feature weights set to zero. But in doing so, I am seeing confusing behavior and -inf lexical reordering feature values.
To perform my experiment, I passed -weight-d 0.011248 0 0 0 0 0 0 on the command line, because my tuned weights had weight-d be a vector of seven nonzero weights, and I wanted to keep the distance penalty the same. When moses starts up, and lists its parameters, I see ... snip weight-d: 0.011248 0 0 0 0 0 0 weight-l: 0.066011 weight-r: -0.040499 weight-t: 0.029361 0.021566 0.067720 0.018145 0.054440 weight-w: -0.146542 snip ... So it looks like I have set the weights to what I want them to be. I expect the best-scoring hypotheses to have zero weights for the lexical reordering features, but instead these features are still mostly nonzero. Furthermore, for a certain percentage of sentences, one feature value becomes -inf, and translation takes greater than 1500 seconds. Here is an example: <example> Translating: 直立 規制 壁 28 の 前 側面 に は 、 更に 、 支持 軸 96 の 左側 に 配置 さ れ た 拘束 片 98 と 支持 軸 96 の 右側 に 配置 さ れ た 受片 100 と が 一体 に 形成 さ れ て いる 。 BEST TRANSLATION: ==== Hypo #10536887, extends #4196134 regulating the front surface of the upright wall 28 further has a support shaft 96 disposed on the left side of a restraining piece 98 and the supporting shaft 96 is arranged on the right side and a receiving member 100 , . have provided out integrally therewith << having >> Translated << が >> [39..39] into << having >> 直立 規制 壁 28 の 前 側面 に は 、 更に 、 支持 軸 96 の 左側 に 配置 さ れ た 拘> 束 片 98 と 支持 軸 96 の 右側 に 配置 さ れ た 受片 100 と << が >> 一体 に 形成 > さ れ て いる 。 Score: [total=nan] <<-44.000, -49.000, 0.000, 1.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, -8.499, -7.371, -10.134, -14.669, -inf, -6.635, -243.237, -49.674, -88.739, -46.117, -90.130, 23.998>> regulating the front surface of the upright wall 28 further has a support shaft 96 disposed on the left side of a restraining piece 98 and the supporting shaft 96 is arranged on the right side and a receiving member 100 , . have provided out integrally therewith having Translation took 1876.630 seconds Finished translating </example> For ease of analysis, these are the excerpted lexical reordering features of the "best translation" with a blazingly bluey nan score: <<-8.499, -7.371, -10.134, -14.669, -inf, -6.635 >> Does anybody know why setting "-weight-d 0.011248 0 0 0 0 0 0" would cause one of the lexical reordering feature weights to go bonkers? Thanks, Jason _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
