Ended 5 years ago
126 participants
236 submissions

Hidden
Nanosemantics Christmas NLP Hackathon

Solve the tasks of chat-based help desk: classify intentions, correct typos, predict responses and fraud

Challenges

Participants are asked to solve the tasks on NLP:

1. To classify the intents

2. To correct the spelling mistakes

3. To predict the scores in the chats

4. To detect the falsification in telephone calls statistics

You have to create algorithms that will solve the classification problem for tasks 1, 3, and 4, and return a corrected sentence for each sentence for task 2.

Quality Criteria

You have to present solutions docker сontainer format. F1-score is the quality metric for the tasks 13 and 4. For the task 2 quality metric is also F1-score, but with an increased penalty factor: the factor is doubled for skipping the error (FN is by a factor of 2), for false detection – increases in 30 times (FP is by a factor of 30).

Prizes

The total prize fund is 500 000 RUB. Teams consisting of up to 4 people of all ages and from all over the world are allowed. The winning teams will be awarded: 

1st place - 250 000 RUB, 

2nd place - 150 000 RUB, 

3rd place - 100 000 RUB. 

The winners are determined by the total metric from all 4 tasks of the competition.

Challenge Description

1. Classification of intents

The intent of the query is the user's intention and the purpose of using a certain search engine. For example, in the case of technical support, classifying the intents can help you automatically cluster queries by their subject to generate a response, or quickly redirect them to the right specialists.

2. Correction of the spelling mistakes

A useful task for your own spell check, and for correcting possible typos of users, for example, when accessing chatbots.

3. Prediction of the estimates in the chats

In this task, we propose to find a correlation between the responses of technical support specialists in chat rooms and their final ratings set by users.

4. Detection of falsification in the telephone calls statistics

Unscrupulous operators sometimes try to falsify the call statistics of their call centers. The key indicator that they try to fake is the duration of the call in seconds. It is suggested to determine the real numbers in reports.

Data

A full description of the data, as well as the data for training models, will be provided to participants at the offline event at Moscow State University.

Submisson Format 

A full description of the format and examples of correct solutions will be provided to participants at the offline event at Moscow State University.

Cookies help us deliver our services. By using our services, you agree to our use of cookies.