CS 440

UIUC AI class

Julie Lovins

SHRDLU (who/what/when)

Who: Terry Winograd

When: 1971

What: early success of natural language model

Zork transcript

Who:

When: 1980

What: Infocom

4 layers of Natural language processing NLP의 4가지 계층 구조

1. semantic (meaning, dialog structure) [

2. Mid-level processing (e.g. part of speech, parsing)

3. low-level text processing (e.g. forming words, finding morphismes)

4. speech [lowest level]

3 tasks of natural language and speech processing

1. convert text/speech into structured information usable by an AI reasoning system (글/말 → 데이터 추출)

2. generate fluent text/speech from structured information (데이터 → 글/말)

3. translate directly between two types of text/speech (between two languages, simple types of question answering) (글/말 → 글/말)

when does modern translation fail

1. uncommon topics

2. less common language

=> the translation system can’t simply regurgitate(되풀이) memorized chunks of text

A classic test to check translation model

Circular translation:

X → Y 번역

Y → X 다시 번역

이 차이 비교

what is ‘spectrogram’?

a visual picture showing how intensity (energy) changes across frequencies over time

(frequency is the y axis / time is x axis / intensity is visualized by color)

how do we know which vowel or semi-vowel it is based on

deep system vs shallow system

1. deep system : includes explicit (명시적) semantic (의미) information (정보)

   1. 의미까지 알아듣기

2. shallow system : operates directly on the low-level or mid-level representations, omitting semantics

   1. 의미(semantics) 단계까지 깊게 가진 않고, 문장 구조 등 하위/중의 수준의 표현상태 (low/mid-level representation)까지만 처리

example of shallow system

words,

word bigrams,

morphemes,

part of speech tags,

shallow parse

example of deep system

Danger of seeming too fluent and the "uncanny valley"

Markov assumption

Assume that only the last few items matter to the next decision.