A kind of caching machine interpretation method and system based on neuronTechnical field
The present invention relates to neuron machine translation mothod.
Background technology
In neuron machine translation field, the statistical machine translation that the decoding time of translation engine is more traditional is much longer.Because the decoding time directly influences the commercialization of machine translation system, therefore, neuron MT engine how is improvedTranslation efficiency is one of most important research topic in neuron machine translation field.At present, to neuron MT engineResearch also rests on the category of decoder, but training pattern and learning framework problem due to neuron MT engine,Even if reconstruction decoder, its decoding efficiency still has the very big rising space.
The content of the invention
Technical problems to be solved first of the invention are to provide a kind of caching machine interpretation method based on neuron,The repetition translation of neuron MT engine can be reduced, improves machine translation server resource utilization and translation efficiency.
The technical proposal for solving the technical problem of the invention is：A kind of caching machine translation based on neuronMethod, comprise the following steps：
（1）User submits the translation comprising cypher text to ask；
（2）The translation request of user is committed to backstage text-processing server and carries out Text Pretreatment by front end Website server；
（3）Cypher text in asking translation is made pauses in reading unpunctuated ancient writings and cutting processing；
（4）Backstage text-processing server is searched in cache whether there is existing translation result, if it is present directlyConnect and translation result is fed back into front end Website server, if it does not exist, then continuing in next step；
（5）Text-processing server in backstage carries out language material standardization to cypher text；
（6）Language material after standardization is committed into machine translation server to be translated, obtains preliminary translation result, and send extremelyBackstage text-processing server；
（7）Text-processing server in backstage reduces to preliminary translation result, generates final translation result；
（8）Text-processing server in backstage sends final translation result to front end Website server, and stores to cache.
Further, Text Pretreatment includes asking included text to carry out the filtering of symbol, mess code translation, makes itText formatting meets the form of the affiliated languages of the text, its punctuation mark is met the punctuation mark mark of the affiliated languages of the textIt is accurate.
Further, punctuate and cutting processing include languages identification and punctuate processing, wherein, languages identification first judges translationThe affiliated languages of text, punctuation mark punctuate rule is made pauses in reading unpunctuated ancient writings and cut to cypher text corresponding to languages belonging to for punctuate processingOffice is managed, and is syntagma by cypher text processing.
Further, language material standardization includes participle and escape.
The present invention also provides a kind of translation system that can run above-mentioned interpretation method, and the system includes front end website serviceDevice, backstage text-processing server, cache and machine translation server, after the output end of front end Website server is connected toThe receiving terminal of platform text server simultaneously is sent to translate solicited message, and the output end of backstage text server is connected to machine and turned overThe receiving terminal and cache of server are translated, and is sent to the translation solicited message by pretreatment, machine translation serverOutput end be connected to the receiving terminal of backstage text server and be sent to preliminary translation result, backstage text server it is defeatedGo out to hold the receiving terminal for being connected to front end Website server and cache and be sent to final translation result.
The beneficial effects of the invention are as follows：The present invention is high by increasing in the Service parts of neuron MT engineFast caching mechanism, when MT engine receives translation request, first it can be scanned for out of cache, if slow at a high speedThere are history translation record, the direct output result of meeting in depositing, it is not necessary to translated by translation engine, also, completed in translation again againAfterwards, translation result and some sentence patterns are introduced directly into cache, it is not necessary to repeat to translate via MT engine, thisInvention greatly improves machine translation server resource utilization and translation efficiency.
Brief description of the drawings
Fig. 1 is the flow chart of the present invention.
Embodiment
Referring to the drawings.
The system of the present invention includes front end Website server, backstage text-processing server, cache and machine translationServer, the output end of front end Website server are connected to the receiving terminal of backstage text server and are sent to translation request letterBreath, the output end of backstage text server is connected to the receiving terminal and cache of machine translation server, and be sent to throughCross the translation solicited message of pretreatment, the output end of machine translation server be connected to the receiving terminal of backstage text server and toIt sends preliminary translation result, and the output end of backstage text server is connected to the receiving terminal of front end Website server and delayed at a high speedDeposit and be sent to final translation result.
The step of being translated using said system is specifically included：
1）User submits the translation comprising cypher text to ask.
Here user refers to the people with machine translation demand, and hardware medium includes page end, the machine of mobile terminal turns overTranslate engine service etc..
2）Cypher text is submitted to backstage text-processing server and carries out Text Pretreatment by front end Website server.
Text Pretreatment refers in text-processing server, and symbol, mess code are carried out by system to text to be translatedFiltering, makes its punctuation mark meet the standard punctuation mark category of the languages.
4）Pretreated cypher text is made pauses in reading unpunctuated ancient writings and cutting is handled.
Punctuate and cutting processing include languages identification and punctuate processing, wherein, languages identification is first judged belonging to cypher textLanguages, punctuate processing punctuation mark punctuate rule corresponding to languages belonging to is made pauses in reading unpunctuated ancient writings by cypher text and cutting processing,It is syntagma by cypher text processing.
5）Backstage text-processing server carries out fast search in cache, search whether existing identical sentenceTranslation result.
1. if it does, directly skipping remaining step, front end Website server is directly returned result to.
2. if it does not exist, then go in next step.
6）The steps such as text-processing server in backstage is segmented to the cypher text after punctuate, escape, such step systemReferred to as language material standardizes.
7）The language material of standardization is submitted into machine translation server to be translated, obtains preliminary translation result.
8）Text-processing server in backstage carries out the operations such as sentence merging, escape reduction to preliminary translation result, and generation is mostWhole translation result.
9）Final translation result is returned to front end Website server by backstage text-processing server, is sent it to simultaneouslyCache is stored.