Tutorial: Artificial Intelligence & Language Processing

What are 3 language processing challenges?

• words never sound the same: different speakers, accents, fast speakers, “lazy'“ speakers, etc

• words are rarely produced without any background noise

• words are rarely produced fullyy

What are some of the ways/cues our minds use to split speech into words and recognise those words?

segmentation cues

• pauses

• stress patterns

• syllable length (longer when part of longer word)

• phonotactics (allowable sound combination)

What are some of the ways/cues our minds use to split speech into words and recognise those words?

flexibility

• adapt to segmentation cues depending on speaker

• adapt to the words a speaker tends to use

What are some of the ways/cues our minds use to split speech into words and recognise those words?

context

• fill in noisy speech signal

• word/sentence context

• situation

• visual cues

What are some of the ways/cues our minds use to split speech into words and recognise those words?

parallel processing

• activate multiple word candidates and modify as the word unfolds

What is automatic speech recognition (ASR) as a system that uses AI to process human language?

• in some cases, it is enough for these systems to just transcribe what we say into text commands

• dictation/transcription

• closed captioning

• call centres

What is Natural Language Processing (NLP) as a system that uses AI to process human language?

• the system needs to not just recognise words but also interpret and respond to our language

• figure out what we’re trying to say and then perform some process/task etc upon that message

What does NLP do?

• processes spoken (or written) language to get at its “meaning”

• transform the message into something the machine can “understand”

• make decisions based on the message

How are NLP systems created?

• historically expert knowledge used to create model of sounds → words

• now, systems are trained to learn relationships between speech/text and the desired response/output

Examples of ASR and NLP in daily life SS

What does AI struggle with (that humans manage)?

• less familiar with different accents

• worse at filling in noisy speech

• less familiar with individual people

• non-literal speech e.g. slang?

What cues/information might AI not use?

• language context

• physical/visual cues

• shared references/knowledge

• diversity of real everyday language

What should future AI take into account?

• non-literal speech (e.g. sarcasm)

• different types of words (e.g. slang words) that are perhaps not part of the “standard training”

• use context better to differentiate between different meanings of ambiguous word

• more diverse input to respond better to different speakers

• input from different languages, including switches between languages

• non-linguistic cues