2020 language and intelligent technology competition was launched, and Baidu provided the largest Chinese data set

                

Language is the most important medium to transmit human information , It is an important challenge to make the machine understand the language and interact with it .3 month 10 Japan ,2020 The competition of language and intelligent technology is officially launched , Open the registration channel for global developers .

This competition is organized by the Chinese information society （CIPS） And China Computer Society （CCF） Co sponsor , Baidu company 、 Evaluation committee of Chinese information society and special committee of Chinese information technology of Chinese computer society jointly undertake , And in the fifth “ Summit Forum on language and intelligence ” Hold technical exchanges and awards , The winning team will share the total amount 35 A bonus of ten thousand yuan . When the , Academic circles at home and abroad 、 Famous experts and scholars in industry , It will also introduce the development trend and innovation achievements of language and intelligence and related fields at home and abroad to the public .

There are five tasks in this competition , Including machine reading comprehension 、 Recommended conversation 、 Semantic analysis 、 Relationship extraction and event extraction , It's about language understanding 、 Man-machine dialogue 、 Knowledge extraction and other complex technologies . Study the above tasks for intelligent search 、 Intelligent recommendation 、 AI applications such as intelligent interaction are of great significance , It is an important frontier topic in the field of natural language processing and artificial intelligence .

The five tasks of this competition will provide Baidu's large-scale Chinese data set , Provide academic exchange platform for researchers , Promote language understanding 、 The development of technology research and application in the field of artificial intelligence .

01

​​ Three classic tasks have been upgraded , Cover more real application scenarios ​​

In this competition , Machine reading comprehension 、 Recommended conversation 、 The three classic tasks of relationship extraction are 2019 We have made a comprehensive upgrade on the basis of .  Machine reading comprehension is to let the machine read the text , Then answer the questions related to the reading . And 2019 Compared to , This year's reading comprehension task , We will focus on the robustness of reading comprehension model in real application scenarios . therefore , In this competition, we specially built DuReader_robust Data sets , Used to examine the robustness of the model in multiple dimensions , Including the hypersensitivity of the model 、 Over stability and generalization ability . The samples in the dataset are all from the actual application scenarios , Difficulty 、 There are many investigation sites , It covers many difficult problems in real applications .  Recommendation oriented dialogue refers to the human-computer interaction system integrating dialogue system and recommendation system , The system first collects users' interests and preferences through questions and answers or chatting , Then actively recommend to users what they are interested in . Real world human-computer interaction involves many kinds of dialogues at the same time , How to naturally integrate multiple types of dialogue is an important challenge .

To meet this challenge , This competition will propose a new task —— Recommendation oriented conversation in multi type conversation . It is expected that the system can actively and naturally lead the conversation from non recommended conversation to recommended conversation , Then based on the collected user interest and user real-time feedback , Complete the final recommendation goal through multiple interactions . meanwhile , Tasks will also provide multiple types of conversations 、 Many fields 、 Integrate users profile Dialog logic data set of information , Close to the real application scenario .  Relation extraction refers to the extraction of entities and their relations from natural language texts . This competition has two upgrades based on last year's information extraction task ：

1. 1. 
      
2. In a simple SPO Based on the relationship, the complex relationship types are added , It is used to depict the complex relations in the real world ;
3. The introduction of Baidu Post Bar oral expression corpus , Its text semantic freedom is higher , More close to the daily oral expression habits , It makes the evaluation task of relation extraction more challenging and practical .
   1. 
      

02

​ Two new hot tasks , A new challenge for the contestants  ​

Different from previous competitions , Besides following machine reading comprehension 、 Recommended conversation 、 Besides the three tasks of relation extraction , In particular, two hot tasks, semantic parsing and event extraction, have been added .  The purpose of semantic parsing task is to enable the machine to automatically convert the natural language problem input by the user into a programming language that can operate with the database （ Such as SQL）, To reduce the threshold and cost of structured data use , At the same time, improve the value and efficiency of the use of structured data .

Current Chinese Text-to-SQL The database of data set is basically made up of single table , The problem model is relatively simple , Only some problems in practical application are covered . This competition will be released for the first time DuSQL Data sets , contain 164 Domain 200 A database ​,​ Covering the match 、 Calculation 、 Reasoning and other common problems in practical applications , Each question is associated with one or more tables in a database . The dataset is closer to the real application scenario , Domain independence for model solving 、 The problem is irrelevant 、 The ability to compute reasoning problems presents a higher challenge .  Event extraction has been widely concerned by academia and industry , It has important practical value , It's also very challenging . In this competition , The goal of the task is to give the target event type and role type set and sentence , Identify all types of events in the sentence , And extract the argument corresponding to the event according to the argument role set . Extract tasks for events , Baidu will release the largest Chinese event extraction data set in the industry , It includes 65 Event types and 1.7 Ten thousand sentences with event information . Hope to pass this competition and open large-scale Chinese data set , Help the further development of event extraction technology . 03

​ Baidu PaddlePaddle fire , Provide full support to the competitors  ​

As the organizer of this competition , Baidu will also provide comprehensive technical resources and platform support for competitors . In this competition , Baidu will provide five competition tasks based on the oar PaddlePaddle Baseline system , Help players quickly get familiar with the competition environment . Open as open source 、 Industry level deep learning platform with complete functions , The flying oar has the core framework of convenient development 、 Support large scale deep learning model training 、 Advanced technologies such as high-performance reasoning engine and industrial open-source model library deployed on multi-terminal and multi platform , Encourage you to use the paddle to complete the design of the model 、 Training and forecasting .  More Than This , Baidu brain AI Studio It will also provide software and hardware environment support for this competition .AI Studio It's a one-stop system based on the paddle platform AI Develop training platform , Provide online programming environment for participating teams 、Tesla V100 free GPU Calculate the force 、 Massive open source algorithms and data . Player login AI Studio You can get the calculation power , Log in every day AI Studio And run Notebook You can get 12 Hourly power , Continuous login 5 Days extra to receive 48 Hourly power .AI Studio Announce that you will sign up for 2020 The language and intelligent technology competition team will provide extra free GPU Calculation time , Completely break the shackles of computing power , Help the players to achieve excellent results .