Through the training on dataset of Automobile's dialogue between customers and the service can we
generate a report automatically.
Overall structure uses Seq2Seq with attention, and embedded
with pre-trained word vector of 100 dim.
The translation uses two solutions of beam search and greedy search.
some example results:
-
Q1:我的帕萨特烧机油怎么办怎么办？,技师说：你好，请问你的车跑了多少公里了，如果在保修期内，可以到当地的4店里面进行检查维修。如果已经超出了保修期建议你到当地的大型维修店进行检查，烧机油一般是发动机活塞环间隙过大和气门油封老化引起的。如果每7500公里烧一升机油的话，可以在后备箱备一些机油，以便机油报警时有机油及时补充，如果超过两升或者两升以上，建议你进行发动机检查维修。|技师说：你好|车主说：嗯,
-
Re:你好 根据 你 的 描述 该车 发动机 烧 机油 如果 超出 了 建议 你 到 4S店 进行 检查 维修
-
Q2:请问 那天右后胎扎了订，补了胎后跑高速80多开始有点抖，110时速以上抖动明显，以为是未做动平衡导致，做了一样抖，请问是不是前面两条胎的问题导致？,技师说：你好师傅！可能前轮平衡快脱落或者不平衡造成的！建议前轮做一下动平衡就好了！希望能够帮到你！祝你用车愉快！|车主说：谢谢大师！|技师说：不客气！祝您用车愉快！
-
Re:你好 根据 你 说 的 这种 情况 建议 你 做 一下 轮胎 动平衡 就 可以 了
some experience:
- Size.
small-sized vocab(3000) with 100dim word vector, 10 encoding units reduce the loss to around 1.1 in 10 epoches , which is very quick. But the result didn't shows good, it can hardly recognized as a semantic sentence. After trying, a vocab of 20k and 256dim word vector, 200 encoding units had a good performance, it is my baseline for this project. - Pretrained word vector.
With pretrained word2vec didn't result in a faster reduction on loss, on the other hand, with trainable embedding layer result in a faster reduction - Batch-size
At a 20k vocab size, 256 word vector dim and 200 hidden units, with 8 batch-size (concerning GPU memory limit) it nearly stopped reduction on loss after 30 epoches(or say a o(10^-3) speed) at around 1.25 and training time for whole dataset costs 1300s. While it had a bounce when increasing to 16 batch-size, loss were about 1.1, and half reduction on training time to 600s, increasing to 32 batch-size caused a little jump as well but not much, while training time stayed at 480s, which wasn't another half of 600.
It shows that if GPU memory supports, bigger batch-size can help the model learn better on data, with faster speed. - Beam search and greedy search
The method is not so much better than greedy search, probably because the training process has always been like a greedy search and it trains pretty well on the ability of deduction on one-to-one word way, instead of a consideration on global optimization of beam search.