Natural Language Processing

Processing Natural Language in a rather intelligent way. To let humans and Agents communicate we need both NLU and NLG.

Natural Language Semantics

• NLP Abstraction
• NLP Ambiguity
• NLP Composition

We call a piece $i$ of information linguistically realized in an utterance $U$, iff, we can trace $i$ to a fragment of $U$.

When an utterance does not realize an information we can still infer it from the utterance or from world knowledge.

![[CleanShot 2023-10-03 at 13.07.22@2x.png]]

![[CleanShot 2023-10-03 at 13.08.34@2x.png]]

Pragmatic Analysis

What helps against wrong semantics is only a restricted domain. This means a restrictied Vocabulary and world model. For example SPARQL on DBPedia.

• NLP Truth Conditions
• NLP Reading

Quantifiers, Scope and Context ![[CleanShot 2023-10-03 at 13.15.25@2x.png]]

Anaphora ![[CleanShot 2023-10-03 at 13.15.58@2x.png]]

Context is personal and keeps changing ![[CleanShot 2023-10-03 at 13.16.20@2x.png]]

• Statistical Language Model

• Text Corpus

• n gram

• N Gram Model (character and word level, see problems with word level)

  • Language Identification
  • Genre Classification
  • Named Entity Recognition

• Out Of Vocabulary

• Perplexity

• POS Tagging

• todo page 72 Viterbi stuff

• Text Classification

• Performance Evaluation Metrics

• Information Retrieval

• Information Extraction

• todo page 108 and before Grammar stuff

• Word Embedding

  • Recurrent Neural Network
  • LSTM

• Sequence To Sequence Model

• Attention

• Greedy Decoding

• Beam Search

• Self-Attention

• Transformer Architecture

• Pretraining

• Transfer Learning

• Masked Language Model

• Language Assistance

• Information Management

  • Search engines
  • Text Classification
  • Information Extraction (QA)

• Dialog Systems

• Topic Modeling

• Sentiment Analysis

• Text Segmentation

  • Named Entity Recognition

• Tokenization

• Stop Word Removal

• Stemming

• Lemmatization

• Cooccurrence

• Large Language Model

  • ChatGPT