Mod6

Terms are _____, Probabilistic, ________-dependent, purposive, and _______.

Subjective, context, evolve over time

Two fundamental obstacles to creating universal term system:

1. Model construction problem

2. 2, symbol grounding

Compositional systems (postcoord) are _____ and _____ to maintain than enumerative systems.

Easier, less expensive

What is coding?

Compression/summarization of data

False Positive (FP)

Assigning a label to pts who DO NOT warrant label

False Negatives (FN)

Failing to assign label to pts who DO warrant label

Reasons for coding errors:

Data errors in pt record (human caused), data not available at time of coding, cannot establish primary purpose of visits (multi-morbidity), coder expertise, data entry errors

Perfect terminology (impossible) has 2 reqs:

1. Ability to cover all concepts

2. Independence from reasoning

Concepts are probabilistic, but ______ are not.

Terminologies

Release cycle for terminologies has two phases:

1. Gradual modification period

2. Major revision

When changing enumerative terminologies, change must be reflected:

Across whole system

When changing compositional terminologies, change only requires:

Change to core terms associated with disease (fewer changes needed than enumerative)

Reference terminologies can be used for:

Automated error checking

Mapping terminologies is easier if they are created from:

Same compositional core (enumerative created independently, compositional created using building block method)

NLP seeks to match _____ with ______.

Context, concepts/terms

Named Entity Recognition (NER)

Process of assigning predefined labels to text (labels may come from terminology or may be names of items)

Word level processing analyzes individual word parts independent of

Linguistic structure or meaning (bag of words)

Tokenization

Converts bag of words into lemma

Stemming

Converting different variants of words into common stems (ex. Catheterization -> catheter)

Lemmatization:

Converts related words into lemma (ex. went -> go)

Regular Expression

Describes in logical form all rules that dictates how a string of symbols can be decomposed

Reference Resolution:

process of using context to determine whether two separate mentions of a concept to same event

Word-Sense Disambiguation

When a word has more than one meaning, selecting the meaning that makes the most sense can only be accomplished by taking into account the context within which the word is used. This concept is known as ________.

Negation

Determining whether a text implies concept "a" or "not a"

N-Grams

Commonly associated words

Parsing

Taking sequent of words and assignign them to different roles in grammar

Query Expansion

Action of a search engine to include synonyms and lexical variants

What is the most common application of NLP?

Information Extraction

Information extraction:

Looks for patterns or performs analysis to locate specific info in text

Information Retrieval:

Keyword searching through tokenization/lemmatization

Used to access documents in large collection

Goal of info retrieval:

Match user's query againt available docs to create list

Question Answering:

User submits NL question, QA system automatically answers in NL

Precise questions —> Direct answers

Advanced text summarization and generation

Text Summarization:

Takes several docs and produces single, coherent text which synthesizes main points

Steps in Summarization Process:

1. Content selection

2. Organization

3. Re-generation

Text Labeling:

Categorization of text into known types

Text Generation:

Formulation of NL sentences from a non-human-readable source

Application Casting:

Deciding which NLP tasks are appropriate to expected output

Topic Modelling:

Takes docs and identifies topics (unsupervised, bag of words approach)

Sequence Labeling:

Leeping track of order in which textual units occur (supervised)

Relation Extraction:

Relation detection and determination of relation type

Approaches to relation extraction:

1. Knowledge-based

2. Statistical

Linguistic Levels:

1. Word-Level (tokens/morphology)

2. Sentence-Level (syntax/semantics)

3. Document-Level (pragmatics/discourse)

Tokens:

Basic language units

Language Units:

Morphemes, words, numbers, symbols, punctuation

Morphology:

words/meaningful parts of words

Combo of morphemes to produce words/lexemes

Morpheme:

root, prefix, suffix

Lexeme:

Forms of the same word (ex. activated, activate, activation)

Syntax:

structure of phrases/sentences

Categorization of words in a language and the structure of phrases/sentences

Discourse structure refers to how author ____.

Organizes info in docs

(aids comprehension for reader)

Coreference Resolution:

Words/phrases referring to same entity

UMLS Metatheseaurs establishes:

Relationships between concepts

Word Embedding

Words/phrases from vocab are numeric vectors

2 Functions of Word Embedding:

1. Capture meaning of word using context

2. Condense representation into vector

Semantics:

Meaning/interpretation of language (words/phrases/sentences)

Entity Linking:

Representing a word with a unique semantic concept

Intrinsic NLP software eval:

Measure changes in system output due to system parameter changes

Extrinsic NLP software eval:

Measure method's performance in a given task

True Positive (TP)

Outputs correctly labeled as HAVING

True Negative (TN)

Outputs correctly labeled as NOT HAVING

Recall (performance assessment)

Number of correct results/gold standard (TP)/(TP+FN)

Precision (performance assessment)

Number of correct results/total results (TP)/(TP+FP)

Clinical Terminologies are important for:

Decision making, PH surveillance, data mining

Pragmatics

Impact of context/intent of speaker on meaning

Discourse

Paragraph/documents

Corpus

Collection of documents

Bag of Words

Word-level processing, analyzes individual words independent of structure/meaning, SIMPLEST way to analyze text

Bag of Words Steps

1. Remove irrelevant text elements (stop words, punctuation)

2. Tokenization

3. Regular expression (stemming, lemmetization)

Many concepts require >1 ______ to describe them.

Lemma

_______ used to present duration in an N-Gram.

Temporal Modifier

In term frequency approach, we prioritize mapping based on:

Historically determine frequency

(Term Frequency) = weight

Inverse Document Frequency (IDF):

Determine weights of rare words in corpus

Parsing

Applying grammar to interpret a texr

Parse Tree

Graphic representation of nested structure of grammar

Grammatical Clues we can use to determine structure of text:

Parts of Speech, Parsing, Context-Tree Grammar, Probabilistic Grammar

Domain knowledge:

Application of ontology, conceptual model of domain, describes all concepts and relationships between concepts

Text Mining

Finding unknown patterns/relationships in clinical texts

Adverse event reporting

Linguistic Stack:

Words - Syntax - Semantics - Discourse - Pragmatics - Formal Understanding - Generation

To extract lab results from Clinical Documents: (Linguistic Stack)

Words - Syntax - Semantics

To extract Pt Hx from Clin Docs: (Linguistic Stack)

Words - Syntax - Semantics - Discourse - Pragmatics - Formal Understanding

To translate pt hx from English to Chinese: (Linguistic Stack)

Words - Syntax - Semantics - Discourse - Pragmatics - Formal Understanding - Generation

To keyword search: (Linguistic Stack)

Words

To do concept search: (Linguistic Stack)

Words - Syntax - Semantics

To do voice-based question answering: (Linguistic Stack)

Words - Syntax - Semantics - Discourse - Pragmatics - Formal Understanding - Generation

Regular Expression Good for:

Simple Pattern Extraction in Word-based application

Regular Expression Bad for:

Anything that needs to understand English

Regular Expression Good for:

Botton-up structure specification, syntactic/semantic parsing

Regular Expression Bad for:

Anything that require a lot of context

Supervised ML Classification requires use to define ____ and _____ but not how they are connected.

Inputs, Outputs

Supervised ML Classification Good for:

A lot of annotated data

Supervised ML Regression similar inputs to __________ by real-valued _____.

Supervised ML Classification, Outputs

Supervised ML Regression good for:

A lot of real-valued data (numeric)

Unsupervised ML Clustering allows for ______ without need for _____.

Grouping similar items, Annotated Data

Unsupervised ML Clustering Good For:

Similarity tasks, NOT needle-in-haystack tasks

Steps to Question Answering:

Step 1: Answer Type Detection (ML Classification)

Step 2: Keyword Extraction (Heuristics)

Step 3: Information Retrieval (Inverted Text)

Step 4: Answer Extraction (ML Classification)

Step 5: Answer Ranking (ML Regression)

Heuristics Method:

Remove stop words, stemming, give high value to terms in UMLS, combine collocations

Inverted Index Method:

Very quick keyword search over large number of docs