时间:2022年9月1日
作者:小芋圆
课程:Special Course in Software Engineering 2022-23
内容:关于该课程的video2的笔记
学习资料、源文件来源于奥卢大学。
1. basic technique on how chatbots understand what the user is trying to tell to the chatbots
two category of user input :
- by text input: keyboard , …
- by speech context : 麦克风,处理器,…
2. spoken/speech as an input modality
this process is called automatic speech recognition (AKA ASR)
- ASR = speech signal processing + pattern recognition
- from speech signal --convert–to–>signal chatbot can understanded
- it’s based on calculate probabilities, rather than key words. 所以没有标准答案(100% correct ),只有更可能的答案
3. ASR process
没什么好说的,如图:
- [speech signal processing]用户说话
- [speech signal processing]麦克风接收
- [speech signal processing]麦克风的信号
- [speech signal processing]被转换成数字信号
- [pattern recognition ]被recognition 为phonemes words
4. an example of what kind of models ASRmight use in converting speech signal into actual language
还没有太看懂
https://ionathan-hui.medium.com/speech-recognition-gmmhmm-8bb5eff8b196
5. challenges/why cannot arrive 100% accuracy
language is very flexible and dynamic .
it’s high context dependence .
if this ASR需要适用于所有人而非特定群体,那么有以下challenges:
其中,第二条speech impediment , i.e. 平翘舌不分,说话漏风,…
6. basic methods of text processing/understand text input
2 methods :
- rule -based approach
- nature language processing (AKA NLP)
7. basic concepts of NLP
as the following 5 concepts :
7.1 tokenization
for Chinese tokenization :(去看文献)
7.2 stemming and lemmatization
tokenization | stemming and lemmatization |
---|---|
分词 | recognition the basic form of the word |
stemming | lemmatization |
---|---|
机械,去掉末尾 | 意译 |
cars -> car | better -> good |
An online stemming and lemmatization tool: snowball
7.3 Part-of-Speech tagging (AKA POS tagging)(划分词性)
i.e.
7.4 chunking / vector semantics (上下文划分)
form words to sentence even to means / intents
通过上下文出现次数的方法,构建坐标系(横纵坐标为两个词),那么每个其他词都可得到坐标(x,y)(x为与横坐标词上下文出现次数,y为与纵坐标词上下文出现次数)
因此,斜率可表示与横纵坐标的相似程度
7.5 named entity recognition
i.e 1
i.e.2
8. summary
- speech recognition --> ASR
- text recognition --> NLP
- tools as examples