Didi open source Delta: AI developers can easily train natural language models

Didi open source DELTA：AI Developers can easily train natural language models

8 month 2 Daily news , Top conference in natural language processing ACL2019 In Florence, Italy . Didi officially announced the open-source training platform for speech and natural language understanding models based on deep learning DELTA, To further help AI Developers create 、 Deploy natural language processing and voice models , Build efficient solutions , help NLP Better application landing .

DELTA It's Dididi 22 Open source projects . There are many natural language processing models and speech models AI The interface between the system and the user , This didi official open source deep learning model training framework , Designed to further reduce developer creation 、 The difficulty of deploying natural language processing systems and voice models .

DELTA Based mainly on TensorFlow structure , To support at the same time NLP( natural language processing ) And speech task and numerical feature training . Including text classification 、 Named entity recognition 、 Natural language reasoning 、 Question and answer 、 Sequence to sequence text generation 、 speech recognition 、 Speaker verification 、 Speech emotion recognition and other important algorithm models , Form a consistent code organization , Unified interface of the whole package .

Users are ready for model training data , And specify the configuration Configuration, model training pipeline Data processing can be done according to the configuration , And choose the corresponding tasks and models , Model training . After training , Automatically generate model file to save . The model file forms a unified interface , It can be used directly online , Fast productization , Can make it easier from research to production .

It is worth noting that , In addition to support multiple models of training ,DELTA It also supports flexible configuration , Developers can be based on DELTA Build dozens of complex models ; Besides ,DELTA It provides stable and efficient benchmark, Users can simply and quickly reproduce the results of the model in the paper , At the same time, we can extend the new model on this basis . After the model is built , Users can use DELTA Deployment process tools for , Get the model online quickly . Seamless connection from paper to product deployment .

at present AI Developers can log in Github(https://github.com/didi/delta) see DELTA Detailed introduction and source code , utilize DELTA Speed up the experiment , Deploy for text categorization 、 Named entity recognition 、 Natural language reasoning 、 Question and answer 、 Sequence to sequence text generation 、 speech recognition 、 Speaker verification 、 Voice emotion recognition and other tasks . Users can also use Didi's open source platform （https://didi.github.io/） Get more information about Didi's open source projects .

actually ,NLP And voice technology has been widely used in didi . Through a large number of applications including natural language processing 、 Deep learning 、 Knowledge map 、 voice 、 Recommended technology , Didi built it on AI Intelligent customer service system , Can use AI technology to assist customer service , Improve the efficiency of manual customer service , And reduce the repetition of manual customer service 、 The amount of processing on simple problems . Besides , Based on speech recognition and natural language understanding technology , Didi is also building driver voice assistant , Didi drivers in Japan and Australia will soon be able to use voice directly “ No contact ” order . And in the future , This voice assistant will also support a full range of voice interaction services , Including video entertainment 、 Information Service 、 Interior environment adjustment , To communicate with passengers 、 Customer service , Even to cheer on 、 Charging or maintenance services . meanwhile , Didi is also actively promoting the opening of relevant capabilities , By providing a one-stop natural language processing tool 、 One stop robot open platform , Help industry partners to better achieve AI Application landing .