Hide Annotations
Beat human teams · of 526
TURNS
0
SUBMISSIONS
0
BEST SCORE
--
TIME
0m
BEST STRATEGY
--

Code Snapshot

waiting
Animation starting...

    

    

      

        
Trajectory

        
Submission 0 / 45 · Turn 0

      

      
    

    

  







  

    

      
Code

      
    

    

      

      

        

        

      

    

  





  

    

      
Trajectory

      
    

    

  





  

    

      
WWW 2025 · Multimodal Dialogue Intent Recognition

      
    

    

      
The WWW 2025 Multimodal Dialogue Intent Recognition competition is jointly organised by Taobao and Tmall Group (淘天集团), the WWW 2025 conference, and the Tianchi platform. It is drawn directly from real e-commerce customer-service operations, where shoppers reach out both by chatting with an agent and by sending screenshots of what they are looking at.

      
Given either a multi-turn Chinese conversation between a customer and a service agent, or a screenshot taken inside a shopping app, the model has to work out what the user actually wants and return a single, fine-grained category label. The text and image inputs are judged together by overall accuracy on a held-out test set, so the real difficulty is labelling consistently and precisely across dozens of closely related categories.


      
Worked examples


      

        
Example 1 · Text dialogue

        
A long thread about wiring and installation; the labelled intent is the single line asking how to control the fan.

        

          

            
INPUT · dialogue

            

              

                
用户：<image>

                
客服：您好！您对这款商品有什么想了解的吗？

                
用户：这款风扇灯安装时对电线有没有特殊要求？

                
客服：需要连接零线、火线和地线；地线没预留可不接，零线火线必须接上。

                
用户：如何操控风扇

                
客服：选好款式颜色后点击图片可看功能参数，支持遥控器操作，对应图片上会有遥控器图标。

                
candidates: 用法用量 · 控制方式 · 能否调光 · 功效功能 · …

              

            

          

          
model
→

          

            
OUTPUT · label

            
控制方式

          

        

      


      

        
Example 2 · Screenshot

        
A checkout page — but the pop-up says the purchase failed, and that is what fixes the label.

        

          

            
INPUT · screenshot

            

              real training-set screenshot
              
Real training-set screenshot (train.json)

            

          

          
model
→

          

            
OUTPUT · label

            
下单过程中出现异常
（显示购买失败浮窗）

          

        

      


      

        
Example 3 · Screenshot

        
Same layout as an ordinary delivery-tracking page; only the “物流异常” banner changes the class.

        

          

            
INPUT · screenshot

            

              real training-set screenshot
              
Real training-set screenshot (train.json)

            

          

          
model
→

          

            
OUTPUT · label

            
物流页面-
物流异常页面

          

        

      


      

        
Example 4 · Screenshot

        
A refund / urge-shipping timeline in which the platform has stepped in to arbitrate the dispute.

        

          

            
INPUT · screenshot

            

              real training-set screenshot
              
Real training-set screenshot (train.json)

            

          

          
model
→

          

            
OUTPUT · label

            
平台介入页面