Is your feature request related to a problem? Please describe.

My dataset is something like the following - just for example:

ID	SrcTag	TgtTag	SrcNoTag	TgtNoTag
okpCtx:tu=54:s=0	<g1>(e) Institutional support</g1>	(e) 机构支持	(e) Institutional support	(e) 机构支持
okpCtx:tu=55:s=0	<g1>(f) Agreeing on the language </g1><x2/><g2>of the mediation</g2>	(f) 就调解语言达成一致	(f) Agreeing on the language of the mediation	(f) 就调解语言达成一致
okpCtx:tu=56:s=0	<g1>1. Commencement of the mediation</g1>	1. 调解的启动	1. Commencement of the mediation	1. 调解的启动
okpCtx:tu=57:s=0	<g1>1. Commencement of the mediation<x2/>5</g1>	<g1>1. 调解的启动<x2/>5</g1>	1. Commencement of the mediation5	1. 调解的启动5
okpCtx:tu=58:s=0	<g1>1. Commencement of the mediation<x2/>6</g1>	<g1>1. 调解的启动<x2/>6</g1>	1. Commencement of the mediation6	1. 调解的启动6
okpCtx:tu=59:s=0	<g1>2. Selection and appointment of a mediator</g1>	2. 调解员的选择和指定	2. Selection and appointment of a mediator	2. 调解员的选择和指定

The SrcTag-TgtTag pairs and SrcNoTag-TgtNoTag pairs are parallel sentences which are translated from English into Chinese. The former have HTML-like tags, while the latter don't have such tags.

Describe the use case

I have been trying to train on the input features of SrcNoTag, TgtNoTag and SrcTag for the output feature of TgtTag, with a config file like the following:

input_features:
    -
        name: SrcTag
        type: sequence
        encoder: 
            type: rnn
            cell_type: lstm
            reduce_output: null
        preprocessing:
          tokenizer: english_tokenize

        name: SrcNoTag
        type: sequence
        encoder: 
            type: rnn
            cell_type: lstm
            reduce_output: null
        preprocessing:
          tokenizer: english_tokenize

        name: TgtNoTag
        type: sequence
        encoder: 
            type: rnn
            cell_type: lstm
            reduce_output: null
        preprocessing:
          tokenizer: chinese_tokenize

output_features:
    -
        name: TgtTag
        type: sequence
        decoder: 
            type: generator
            cell_type: lstm
            attention: bahdanau
            reduce_input: null
        #loss:
            #type: softmax_cross_entropy
        preprocessing:
          tokenizer: chinese_tokenize

training:
    batch_size: 8

Describe the solution you'd like

The above trained sucessfully, but does not really give any meaningful result.

Describe alternatives you've considered

I have considered a "combiner", but do not know how.

Additional context

As you may guess, the model is supposed to re-construct the tags that are in the source English sentences in the tag-free target Chinese sentences.
I have tried the very first record on You Chat - https://you.com/search?q=best+laptops&fromSearchBar=true&tbm=youchat , and got a very promising result:

(Though it wrongly repeated the result as its subject.)