Hello, I'm very happy to use such a great service. I have some questions to ask, please feel free to enlighten me .
1. Is there a "confidence" data for recognition results and candidate words?
2. Currently, the client can generate LK resources through RecognitionAssetsBuilder, and LK + SK resources can be generated through online tools. Is there a way for the client or our server to generate SK resources?
3. AK files now have lite and regular versions. Are there versions that are trained separately for specific fields? In our business scenarios, the range of user handwritten content is relatively fixed.
4. LK and SK are both whitelist modes. Is there a blacklist mechanism? In our business scenarios, some characters will definitely not appear, which should also improve the accuracy.
5. We found that it is easy to make mistakes when inputting "numbers + Chinese characters", for example, "3斤" is recognized as "沂" "1包" is recognized as "泡" Is there any best practice in this area?
Kongyan4696
你好，很高兴能使用到这么棒的服务，有些问题想咨询下，请不吝赐教
- Oldest First
- Popular
- Newest First
Sorted by PopularEtienne.dauphin
If you do have precise questions, please proceed.
1 person likes this
Kongyan4696
--- 帖子内容带有表情，导致所有文字消失了 :(
你好，很高兴能使用到这么棒的服务，有些问题想咨询下，请不吝赐教。
1. 识别结果和候选词有没有 "置信度" 这个数据？
2. 目前客户端能通过 RecognitionAssetsBuilder 生成 LK 资源，通过在线工具能生成 LK + SK 资源。 请问客户端或者我们服务器有办法生成 SK 资源吗？
3. AK 文件现在有lite版和普通版，有没有细分领域单独训练的版本？我们业务场景下，用户手写的内容范围比较固定。
4. LK、SK 都是白名单模式，有没有黑名单机制？在我们业务场景下某些文字肯定是不会出现的，这样应该也能提升准确率。
5. 我们发现 "数字+汉字" 的输入下比较容易出错，比如
"3斤" 识别成了 "沂"
"1包" 识别成了 "泡"
请问下这块有最佳实践吗？
---translate---
Hello, I'm very happy to use such a great service. I have some questions to ask, please feel free to enlighten me .
1. Is there a "confidence" data for recognition results and candidate words?
2. Currently, the client can generate LK resources through RecognitionAssetsBuilder, and LK + SK resources can be generated through online tools. Is there a way for the client or our server to generate SK resources?
3. AK files now have lite and regular versions. Are there versions that are trained separately for specific fields? In our business scenarios, the range of user handwritten content is relatively fixed.
4. LK and SK are both whitelist modes. Is there a blacklist mechanism? In our business scenarios, some characters will definitely not appear, which should also improve the accuracy.
5. We found that it is easy to make mistakes when inputting "numbers + Chinese characters", for example,
"3斤" is recognized as "沂"
"1包" is recognized as "泡"
Is there any best practice in this area?
Kongyan4696
:)