# 問答功能
###### tags: `Speaker` `Function`
{%hackmd theme-dark %}
[HomePage](https://hackmd.io/Irmd-eDAT_2WT__nWnpNTg)
[音箱](https://hackmd.io/sFQWkFArTxSiRt2AorTGCQ)

## Keyword
- "幫我查" + question string
- any queries that can't be handled by other functions
## Input
- question string
## output
- answer string (could be empty)

## description
This is a function for searching answers for questions by scraping websites(e.g. wiki).
when no question found, output is default to "".
If there is multiple possible answers, answers will be seperated by '/'.

## 流程
1. 接收查詢字串
2. 透過網路搜尋相關資料 / 使用現成API搜尋資料
~~3. 使用transformer、QA model從網路資料中擷取答案 / 使用現成API擷取資料~~
4. 輸出可能的答案字串 (提供多組答案，但不一定皆為正解，會提供出處)

## API
External API
<!-- - [English Wikipedia API](https://en.wikipedia.org/w/api.php)
- [Google Custom search](https://developers.google.com/custom-search/v1/overview) (100 free query per day) -->
~~- [duckduckgo](https://duckduckgo.com/api)~~
~~- [hugging face](https://huggingface.co/transformers/pretrained_models.html)~~
Python package
- transformers (require python 3.6 above)
- beautifulsoup
- pytorch

## Resources
- Internet access

## Todo
~~[x] Check google API~~
- [ ] Setup test system
- [ ] Chinese support
- [ ] Passage Retriever (with duckduckgo)
- [ ] Chinese model (BERT trained with chinese)

## Problems
~~- Preferable API?~~
~~- Use existing API / make one using other API /~~ scrape internet for answer
- Model to use for IR? (affects accuracy and speed)
~~- Where to obtain answers? (Google / Wiki)~~

## Reference
### build a QA system from scratch
<!-- - [spaCy](https://spacy.io/usage/spacy-101)
- [Media Wiki API](https://www.mediawiki.org/wiki/API:Main_page) -->
<!-- - [haystack](https://github.com/deepset-ai/haystack) -->
<!-- - [Google search serpAPI](https://serpapi.com/search-api) -->
- [QA from scratch](https://towardsdatascience.com/building-an-application-of-question-answering-system-from-scratch-2dfc53f760aa)
- [Transformer QA in another language](https://zhuanlan.zhihu.com/p/345589230)
<!-- - [Riva QA and Wiki API](https://developer.nvidia.com/blog/developing-a-question-answering-application-quickly-using-riva/) -->
<!-- - [Question Answering With Spokestack and Transformers](https://www.spokestack.io/blog/building-a-question-answering-bot-with-python) -->
<!-- - [BERT 簡易 QA](https://zhuanlan.zhihu.com/p/82391263) -->
- [Wiki with BERT](https://qa.fastforwardlabs.com/pytorch/hugging%20face/wikipedia/bert/transformers/2020/05/19/Getting_Started_with_QA.html)
<!-- - [facebook sentence embedding](https://towardsdatascience.com/building-a-question-answering-system-part-1-9388aadff507) -->
- [BERT Question Answering in TensorFlow](https://medium.com/nvidia-ai/how-to-train-bert-from-scratch-on-gpus-a9603b0cb60e)

### use exsiting API
- [WolframAlpha](https://products.wolframalpha.com/conversational-api/documentation/)

<!--以下為回饋＆評論區-->
### 測試組提供的使用情境
註：Noname為音箱名稱
（指令制式化＜幫我查/請去找天氣/新聞＞或口語化關鍵字搜尋）

hey Noname 今天天氣如何 ->（google氣象）你所在的地區……
hey Noname 他今天生什麼氣阿->（不明確指令）我不清楚您……
hey Noname 有發生什麼事嗎->（google新聞）根據……發生xxx
hey Noname 什麼是xxx->（wiki）xxx是指在A國……
hey Noname 那A國的匯率怎麼樣->（不明確指令）
->（關鍵字搜尋）由…網站提供……
hey Noname 幫我查A國股市->（google財經）根據標準時間……