final hpsgs cleaning up and final aspects, semantics, overview to statistical nlp

CSE6339 3.0 Introduction to Computational LinguisticsTuesdays, Thursdays 14:30-16:00 – South Ross 101

Fall Semester, 2011

Instructor: Nick Cercone - 3050 CSEB - [email protected] 1

Final HPSGs

Cleaning up and final aspects, semantics, overview to statistical NLP


Fall Semester, 2011


HPSGs

An Overlooked Topic: Complements vs. Modifiers• Intuitive idea: Complements introduce essential

participants in the situation denoted; modifiers refine the description.

• Generally accepted distinction, but disputes over individual cases.

• Linguists rely on heuristics to decide how to analyze questionable cases (usually PPs).


Fall Semester, 2011


HPSGs

Heuristics for Complements vs. Modifiers

• Obligatory PPs are usually complements.

• Temporal & locative PPs are usually modifiers.

• An entailment test: If X Ved (NP) PP does not entail X did something PP, then the PP is a complement.

Examples

– Pat relied on Chris does not entail Pat did something on Chris

– Pat put nuts in a cup does not entail Pat did something in a cup

– Pat slept until noon does entail Pat did something until noon

– Pat ate lunch at Bytes does entail Pat did something at Bytes


Fall Semester, 2011


HPSGs

Agreement

• Two kinds so far (namely?)

• Both initially handled via stipulation in theHead-Specifier Rule

• But if we want to use this rule for categories that don’t have the AGR feature (such as PPs and APs, in English), we can’t build it into the rule.


Fall Semester, 2011


HPSGs

The Specifier-Head Agreement Constraint (SHAC)

Verbs and nouns must be specified as:


Fall Semester, 2011


HPSGsThe Count/Mass Distinction

• Partially semantically motivated– mass terms tend to refer to undifferentiated substances (air, butter,

courtesy, information)

– count nouns tend to refer to individuatable entities (bird, cookie, insult, fact)

• But there are exceptions:– succotash (mass) denotes a mix of corn & lima beans, so it’s not

undifferentiated.

– furniture, footwear, cutlery, etc. refer to individuatable artifacts with mass terms

– cabbage can be either count or mass, but many speakers get lettuce only as mass.


Fall Semester, 2011


HPSGs – Semantics

The Linguist’s stance: Building a precise model• Some statements are statements about how the model works:

“[prep] and [AGR 3sing] cannot be combined because AGR is not a feature of the type prep.”

• Some statements are statements about how (we think) English or language in general works.

“The determiners a and many only occur with count nouns, the determiner

much only occurs with mass nouns, and the determiner the occurs with either.”

• Some are statements about how we code a particular linguistic fact within the model.

“All count nouns are [SPR < [COUNT +]>].”


Fall Semester, 2011


HPSGs – Semantics

The Linguist’s stance:A Vista on the Set of Possible English Sentences

• ... as a background against which linguistic elements (words, phrases) have a distribution

• ... as an arena in which linguistic elements “behave” in certain ways


Fall Semester, 2011


HPSGs - Semantics

So far, our “grammar” has no semantic representations. We have, however, been relying on semantic intuitions in our argumentation, and discussing semantic contrasts where they line up (or don't) with syntactic ones.

Examples?• structural ambiguity

• S/NP parallelism

• count/mass distinction

• complements vs. modifiers


Fall Semester, 2011


HPSGs - Semantics

Aspects of meaning we won’t account for

• Pragmatics

• Fine-grained lexical semantics:


Fall Semester, 2011


HPSGs - Semantics

Our Slice of a World of Meanings

“... the linguistic meaning of Chris saved Pat is a proposition that will be true just in case there is an actual situation that involves the saving of someone named Pat by someone named Chris.”


Fall Semester, 2011


HPSGs - Semantics

Our Slice of a World of Meanings

What we are accounting for is the compositionality of

sentence meaning.

• How the pieces fit togetherSemantic arguments and indices

• How the meanings of the parts add up to the meaning of the whole.

Appending RESTR lists up the tree


Fall Semester, 2011


HPSGs – Semantics in Constraint-based grammar

Constraints as generalized truth conditions• proposition: what must be the case for a proposition to be true

• directive: what must happen for a directive to be fulfilled

• question: the kind of situation the asker is asking about

• reference: the kind of entity the speaker is referring to

Syntax/semantics interface: Constraints on how syntactic arguments are related to semantic ones, and on how semantic information is compiled from different parts of the sentence.


Fall Semester, 2011


HPSGs – Semantics – Feature Geometry


Fall Semester, 2011


HPSGs – Semantics – How the pieces fit together


Fall Semester, 2011




Fall Semester, 2011


HPSGs – Semantics (pieces together)


Fall Semester, 2011


HPSGs – Semantics (more detailed view of same tree)


Fall Semester, 2011


HPSGs – Semantics

To Fill in Semantics for the S-node, we need the Semantics Principles

The Semantic Inheritance Principle:• In any headed phrase, the mother's MODE and INDEX are

identical to those of the head daughter.

The Semantic Compositionality Principle:• In any well-formed phrase structure, the mother's RESTR

value is the sum of the RESTR values of the daughter.


Fall Semester, 2011


HPSGs – Semantics – semantics inheritance illustrated


Fall Semester, 2011


HPSGs – Semantics - semantic compositionality illustrated


Fall Semester, 2011


HPSGs – Semantics – what identifies indices


Fall Semester, 2011


HPSGs – Semantics – summary wordscontribute predications

‘expose’ one index in those predications, for use by words or phrases

relate syntactic arguments to semantic arguments


Fall Semester, 2011


HPSGs – Semantics – summary, grammar rules

identify feature structures (including the INDEX value) across daughters

Head Specifier

Rule

Head Complement

Rule

Head Modifier

Rule


Fall Semester, 2011


HPSGs – Semantics – summary, grammar rulesidentify feature structures (including the INDEX value) across daughters

license trees which are subject to the semantic principles- SIP ‘passes up’ MODE and INDEX from head daughter


Fall Semester, 2011


HPSGs – Semantics – summary, grammar rulesidentify feature structures (including the INDEX value) across daughters

license trees which are subject to the semantic principles

- SIP ‘passes up’ MODE and INDEX from head daughter

- SCP: ‘gathers up’ predications (RESTR list) from all daughters


Fall Semester, 2011


HPSGs – other aspects of semantics

• Tense, Quantification (only touched on here)• Modification• Coordination• Structural Ambiguity


Fall Semester, 2011


HPSGs – what were are trying to do

Objectives

• Develop a theory of knowledge of language

• Represent linguistic information explicitly enough to distinguish well-formed from ill-formed expressions

• Be parsimonious, capturing linguistically significant generalizations.

Why Formalize?

• To formulate testable predictions

• To check for consistency

• To make it possible to get a computer to do it for us


Fall Semester, 2011


HPSGs –how we construct sentences

The Components of Our Grammar• Grammar rules

• Lexical entries

• Principles

• Type hierarchy (very preliminary, so far)

• Initial symbol (S, for now)

We combine constraints from these components.• Question: What says we have to combine them?


Fall Semester, 2011


HPSGs – an example

A cat slept.

• Can we build this with our tools?

• Given the constraints our grammar puts on well-formed sentences, is this one?


Fall Semester, 2011


HPSGs – lexical entry for “a”

• Is this a fully specified description?

• What features are unspecified? • How many word structures can this entry license?


Fall Semester, 2011


HPSGs – lexical entry for “cat”

• Which feature paths are abbreviated and Is this fully specified?

• What features are unspecified? • How many word structures can this entry license?


Fall Semester, 2011


HPSGs - Effect of Principles: the SHAC


Fall Semester, 2011


HPSGs - Description of Word Structures for cat


Fall Semester, 2011


HPSGs - Description of Word Structures for a


Fall Semester, 2011


HPSGs - Building a Phrase


Fall Semester, 2011


HPSGs - Constraints Contributed by Daughter Subtrees


Fall Semester, 2011


HPSGs - Constraints Contributed by the Grammar Rule


Fall Semester, 2011


HPSGs - A Constraint Involving the SHAC


Fall Semester, 2011


HPSGs - Effects of the Valence Principle


Fall Semester, 2011


HPSGs - Effects of the Head Feature Principle


Fall Semester, 2011


HPSGs - Effects of the Semantic Inheritance Principle


Fall Semester, 2011


HPSGs - Effects of the Semantic Compositionality Principle


Fall Semester, 2011


HPSGs - Is the Mother Node Now Completely Specified?


Fall Semester, 2011


HPSGs - Lexical Entry for slept


Fall Semester, 2011


HPSGs - Another Head-Specifier Phrase


Fall Semester, 2011


HPSGs - Is this description fully specified?


Fall Semester, 2011


HPSGs - Does the top node satisfy the initial symbol?


Fall Semester, 2011


HPSGs - RESTR of the S node


Fall Semester, 2011


HPSGs – Another example


Fall Semester, 2011


HPSGs - Head Features from Lexical Entries


Fall Semester, 2011


HPSGs - Head Features from Lexical Entries, plus HFP


Fall Semester, 2011


HPSGs - Valence Features:Lexicon, Rules, and the Valence Principle


Fall Semester, 2011


HPSGs - Required Identities: Grammar Rules


Fall Semester, 2011


HPSGs - Two Semantic Features: the Lexicon & SIP


Fall Semester, 2011


HPSGs - RESTR Values and the SCP


Fall Semester, 2011


HPSGs - An Ungrammatical Example

What’s wrong with this sentence?

The Valence Principle, Head Specifier Rule


Fall Semester, 2011


HPSGs – Overview

• Information movement in trees

• Exercise in critical thinking

• SPR and COMPS

• Technical details (lexical entries, trees)

• Analogies to other systems you might know, e.g., How is the type hierarchy like an ontology?


Fall Semester, 2011


Statistical NLP – Introduction

NLP as we have examined thus far can be contrasted with statistical NLP. For example, statistical parsing researchers assue that there is a continuum and that the only distinction to be drawn is between the correct parse and all the rest.

The “parse” given by the parse tree on the

right would support this continuum view.

For statistical NLP researchers, there is no

Difference between parsing and syntactic

Disambiguation: its parsing all the way!


Fall Semester, 2011


Statistical NLP – Statistical NLP is normally taught in 2 parts:• Part I lays out the mathematical and linguistic foundation that the other

parts build on. These include concepts and techniques normally referred to throughout the course.

• Part II covers word-centered work in Statistical NLP. There is a natural progression from simple to complex linguistic phenomena in collocations, n-gram models, word sense disambiguation, and lexical acquisition. This work is followed by techniques such as Markov Models, tagging, probabilistic context free grammars, and probabilistic parsing, which build on each other. Finally other applications and techniques are introduced: statistical alignment and machine translation, clustering, information retrieval, and text categorization.


Fall Semester, 2011


Statistical NLP – What we will discuss

1. Information Retrieval and the Vector Space Model Typical IR system architecture, steps in document and query processing in IR, vector space model, tfidf - term frequency inverse document frequency weights, term weighting formula, cosine similarity measure, term-by-document matrix, reducing the number of dimensions, Latent Semantic Analysis, IR evaluation


Fall Semester, 2011


Statistical NLP - – What we will discuss

2. Text Classification Text classification and text clustering, Types of text classification, evaluation measures in text classification, F-measure, Evaluation methods for classification: general issues - over fitting and under fitting, methods: 1. training error, 2. train and test, 3. n-fold cross-validation


Fall Semester, 2011



3. Parser Evaluation, Text Clustering and CNG Classification Parser evaluation: PARSEVAL measures, labeled and unlabeled precision and recall, F-measure; Text clustering: task definition, the simple k-means method, hierarchical clustering, divisive and agglomerative clustering; evaluation of clustering: inter-cluster similarity, cluster purity, use of entropy or information gain; CNG -- Common N-Grams classification method


Fall Semester, 2011



4. Probabilistic Modeling and Joint Distribution ModelElements of probability theory, Generative models, Bayesian inference, Probabilistic modeling: random variables, random configurations, computational tasks in probabilistic modeling, spam detection example, joint distribution model, drawbacks of joint distribution model


Fall Semester, 2011



5. Fully Independent Model and Naive Bayes ModelFully independent model, example, computational tasks, sum-product formula; Naive Bayes model: motivation, assumption, computational tasks, example, number of parameters, pros and cons; N-gram model, language modeling in speech recognition


Fall Semester, 2011



6. N-gram Model N-gram model: n-gram model assumption, graphical representation, use of log probabilities; Markov chain: stochastic process, Markov process, Markov chain; Perplexity and evaluation of N-gram models, Text classification using language models


Fall Semester, 2011



7. Hidden Markov Model Smoothing: Add-one (Laplace) smoothing, Bell-Witten smoothing; Hidden Markov Model, graphical representations, assumption, HMM POS example, Viterbi algorithm -- use of dynamic programming in HMMs.


Fall Semester, 2011



8. Bayesian Networks Bayesian Networks, definition, example, Evaluation tasks in Bayesian Networks: evaluation, sampling, inference in Bayesian Networks by brute force, general inference in Bayesian Networks is NP-hard, efficient inference in Bayesian Networks,


Fall Semester, 2011


Other Concluding Remarks

ATOMYRIADES

Nature, it seems, is the popular namefor milliards and milliards and milliardsof particles playing their infinite gameof billiards and billiards and billiards.

final hpsgs cleaning up and final aspects, semantics, overview to statistical nlp

Documents