Teadmiste formaliseerimine

teadmised

Name: Knowlege representation Code: ITI8700 Link: http://lambda.ee/Teadmiste_formaliseerimine Lecturer: Tanel Tammet Practice work: Priit Järv Contact: tanel.tammet@ttu.ee, 6203457, TTÜ ICT-426 Archives of previous years: 2019.

Time, place, result

Lectures: Friday 10:15-11:45 room ICT-315
Practical work: Friday 12:00-13:30 room ICT-315
Practical work will give 40% and exam 60% of points underlying the final grade. The exam will consist of several small excercises.

The first practical work time on 31. January at 12:00 will be used for introducin the practice work.

Some of the last lecture times at the end of the course may be used as additional seminars/labs.

Focus

The main focus of the course is on knowledge representation and reasoning (KR), on a spectrum from simple (representing and using knowledge in databases) to complex (meaning of sentences in natural language and commonsense knowledge).

The overall goal of the labwork to understand issues facing a task of building a natural language question answering system a la Watson and to see and experiment with existing parts of the solution.

The course contains the following blocks:

• Databasis
• Representing and using simple facts and rules in symbolic AI and logic.
• General-knowledge databases: wikidata, wordnet, yago, conceptnet, nell, cyc.
• Knowledge conveyed by natural sentences: both statistical methods like word2vec and logic-based methods.
• Reasoners and question answering systems.
• Confidence and relevance.
• Context, time, location, events, causality.

Check out this description of the whole KR area.

Books to use

• Recommended background for NLP: Speech and Language Processing by Dan Jurafsky and James H. Martin. here is the 2nd edition and here the web page for the draft 3rd edition
• Ca half of the course themes are covered in this book with freely accessible pdf-s.
• The freely accesible pdf of the handbook if knowledge representation gives a detailed and thorough coverage of the subject; far more than necessary for the course.

Observe that a noticeable part of the course contents are not covered by these books: use the course materials and links to papers, standards and tutorials provided.

Practical work

There are three labs. They are all steps in a single project: build a simple natural-language question-answering system.

The labs have to be presented to the course teachers and all students present at labwork time.

The labs can be prepared alone or by teams of two people.

First, read the explanation of the overall task for the recent 2019 course. The contents will me modified and updated for 2020. The following labs are steps on the path to the end goal.

First lab

Please read about the task for the first lab

Second lab

The task in the second lab is actually answering simple questions using logic and a small set of rules you write and a reasoner.

Please start by experimenting with a real prover: look and run Reasoner examples with gkc

We use the gkc prover for examples (instructions are in the Reasoner examples ... page above) but you could also try out an old classic prover Otter which actually works OK on windows even though it is really old.

Make sure to get the latest release of gkc from gkc releases (currently delta) and either download a binary or compile using instructions on the gkc github page. The basic instructions are also on the gkc github page. Also, see the few examples in the Examples folder.

Third lab

Add an NLP component to the second lab to answer questions posed in natural language, not logic.

You can use a trivial NLP-to-reasoner parser in python nlp.py as a starting point.

Have a look at the Lecture 12 materials.

Interesting large question-answering datasets and challenge problems: Stanford SQUAD, Allen institute arc2 challenge, Google natural questions, Fujitsu NLP challenge, amazon qa dataset

Lecture block 1: basics and representing simple facts.

Lectures 1 and 2: Overview of the course. Background and basics: SQL, logic, NLP

Lecture materials:

• intro, sql, logic, NLP

Lecture 2: RDF and RDFS (and OWL)

Lecture materials:

• schemaless databases, RDF and RDFS
• Nosql links and notes

Lecture block 2: capturing meaning in natural language

Lecture 3: Intro to homework and NLP

Lecture materials:

• nlp and watson
• Jurafsky and Martin book: conversational agents
• start reading the Jurafsky and Martin book (see links above)

Lecture 4: vector representation of words

This lecture will be given by Priit Järv.

Lecture materials:

• First, listen to text vectorization episode from the linear digressions podcast.
• Second, listen to word2vec episode from the linear digressions podcast.
• Read the beginning (until you get tired ...) of the Vector Semantics chapter from the Jurafsky & Martin book.

Useful additional materials from (roughly) easier to more complex:

• word vectors: a short intro
• vector semantics part I presentation from the Jurafsky & Martin book for for chapter 6.
• vector representation: a short tutorial with code examples.
• word2vec tutorial from tensorflow
• vector semantics part II presentation from the Jurafsky & Martin book for chapter 6.

Probabilistic models:

• LDA explanation by Jordan Boyd-Graber

Also interesting to read:

• Aetherial Symbols by the ultimate machine learning guru aka "godfather of machine learning" Geoffrey Hinton
• Autocomplete using Markov chains

Lecture block 3 and start of 4: large common-sense knowledge bases and reasoning

Lecture 5: First look into main large knowledge bases

We will have a look at the goals, main content and differences between:

• wordnet
• dbpedia
• yago
• babelnet
• conceptnet
• nell in wiki and a good paper. The nell website is currently down.
• cyc

Lecture 6: Big annotation systems and intro to rule reasoners

First, big annotation systems:

• Google structured data on the webpage.
• Facebook Open Graph markup on the webpage.

In connection, see also:

• schema.org: property markup vocabulary suggested by Google, Microsoft and others.
• json-ld: currently most popular rdf syntax (in json), also recommended by Google.

Second, intro to rule reasoners:

• Lecture material as ppt or as pdf

Additionally you may want to look at:

• automated reasoning tutorial by Geoff

Lecture 7: rule reasoners

Please have a look and run the same experiments yourself with the gkc prover:

Reasoner examples with gkc

We use the gkc prover for examples (instructions are in the Reasoner examples ... page above) but you could also try out an old classic prover Otter which actually works OK on windows even though it is really old.

Lecture 8: starting to parse natural language for interaction with rule reasoners

We will start building a tiny question-answering tool, translating natural language to rule reasoner input.

Our first, very simple goal: translate the text "John is a father of Pete. Pete is a father of Mark. Who is the father of Pete?" to a rule reasoner input

father(john,pete).
father(pete,mark).
-father(X,pete) | ans(X).

then run the reasoner, fetch the ans(john) and give an answer "John is".

You can use a trivial NLP-to-reasoner parser in python nlp.py as a starting point.

We will then proceed to gradually make questions and sentences more complex and add rules to be able to answer them.

In the final part of the course we will start looking at uncertain reasoning: right now we will only try to do simple strict reasoning.

We will have a look at potential tools for semantic parsing, from simpler to more complex:

• parsetron: a really simple and very limited tool to build your own parser.
• sippycup: has a detailed tutorial/course attached, check the first lab
• ccg2lambda: parses to classical logic strings, which are embedded in XML output.

Other interesting things to look at:

• Framenet: map meaning to form in English through the theory of Frame Semantics.
• Sling: Google project for getting facts from Wikipedia using Frame Semantics.
• Spider: Yale Semantic Parsing and Text-to-SQL Challenge.
• WikiSQL: A Salesforce crowd-sourced dataset for developing NLP interfaces for relational databases.
• Allennlp from the Allen Institute for AI. See also their semantic parsing tutorial.

Lecture block 5: context and uncertainty

Lecture 9: intro to representing context and uncertainty

Blocks world.

• Old intro: planning and blocks world

Some axiom sets:

• PLA001-0.ax
• PLA001-1.ax

And some planning problems to solve from these axiom sets:

• PLA004-1.p
• PLA005-1.p
• PLA019-1.p

Also:

Alternative axiomatization

Next, starting uncertainty.

We will first look at default logic.

Lecture 10: various classical ideas for discrete reasoning with uncertainty

We will look at some discrete ways:

• default logic: refresher from previous lecture.
• circumscription
• Autoepistemic logic: long tutorial Ijcai93.pdf

And start with numeric:

• Uncertain_prob_fuzzy.ppt‎ Intro to Bayesian and fuzzy reasoning.
• Notes about numerical confidence, popularity, etc measures

Lecture 11: various classical ideas for numeric reasoning with uncertainty

Classical numeric ways to reason about uncertainty:

• Uncertain_prob_fuzzy.ppt‎ Intro to both Bayesian and Fuzzy reasoning.

Some newer approaches:

• Vienna_tanel_4.pdf Additional examples and combining.
• Problog start with a tutorial

Lecture 12: seminar-kind of work on the second lab

Background to start with:

• Take and unpack gkc.zip from gkc release gamma

• Start with the example file p1.txt

father(john,pete).

% either the first or a second or both facts are true:
father(pete,mark).

% equivalent to (father(X,Y) & father(Y,Z)) => grandfather(X,Z).
-father(X,Y) | -father(Y,Z) | grandfather(X,Z).

-grandfather(john,X) | ans(X).

• Run

  gkc p1.txt

• Start with the trivial NLP-to-reasoner parser in python nlp.py .

• Our initial goal is to process the text "John is a father of Andrew. Andrew is a man. Who is the father of Andrew?"

• Reasoner example created by nlp.py
• Reasoner example with some rules for relatives

And the examples with various ways to encode normal/abnormal cases (not really needed for lab):

• Reasoner example for being likely married: works out ok
• Reasoner example for birds: does not work out well
• Reasoner another example for birds: does not work out well

Syntax and command line:

• gkc command line: gkc readme in github: see also Examples folder for a few trivial examples.
• For simple clausal syntax see otter manual clause syntax.
• For complex syntax options see TPTP manual: seems to be sometimes down.

Lecture 13: seminar-kind of work on the third lab

Ways to get external data in addition to our own rules and data:

• Raw data (hard to use): wikipedia, wikidata
• Processed raw data (a bit easier): dbpedia (processed wikipedia/wikidata)
• Well-formed and connected options (use one of these):
  • wordnet: exists in several versions, including an external TPTP version usable by the gkc reasoner, direct link to TPTP axiom NLP001+0.ax
  • yago, connecting wikipedia with wordnet and geonames. Look at the downloads section SIMPLETAX and download yagoSimpleTypes and yagoSimpleTaxonomy. You will need to write a program to filter and convert the yago datasets to the prover-understandable form. See some ideas for converting yago.

See how to run gkc on the Wordnet TPTP version