Presentation is loading. Please wait.

Presentation is loading. Please wait.

What‘s an Event ? How Ontologies and Linguistic Semantics Shape Upcoming Challenges for Machine Learning Models of Event Extraction Udo Hahn Jena University.

Similar presentations


Presentation on theme: "What‘s an Event ? How Ontologies and Linguistic Semantics Shape Upcoming Challenges for Machine Learning Models of Event Extraction Udo Hahn Jena University."— Presentation transcript:

1 What‘s an Event ? How Ontologies and Linguistic Semantics Shape Upcoming Challenges for Machine Learning Models of Event Extraction Udo Hahn Jena University Language and Information Engineering (JULIE) Lab ECML PKDD 2008 European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases Workshop on High-Level Information Extraction September, Antwerp, Belgium

2 2 Relations, Events, Semantic Roles,.. Narrative Report: John Doe is a naturalized United States citizen from Iran. His immediate family (father, mother and two sisters) still live in Iran. … He has traveled to Gaza 6x w/in the last year. John Doe: [category = person] citizenship = United States origin = Iran relative = father … father: [category = person] residence = Iran event = traveled arg0 = he (John Doe) arg1= Gaza frequency = 6 times duration = the last year ACE TDT SemEval

3 3 Relational Knowledge Types Representation of Static Knowledge („invariable“ states, are there any ??? ) –Conceptual: Is-a, Instance-of –Conceptual: Part-of –Attributive (properties, binary relations): located-in, citizen-of, has-socialsecurityno,... Representation of Dynamic Knowledge (state changes, dependencies among states) –processes, actions, events sales, rating changes, transports, launches, mergers & acquisitions, kidnappings, …

4 4 Hand Washing as an Event

5 5 Hand Washing Event (1/2) Hand washing: A simple way to prevent infection Hand washing is a simple habit that can help keep you healthy. Learn about the benefits of good hand hygiene, as well as when to wash your hands and how to clean them properly. Hand washing is a simple habit — one that requires minimal training and no special equipment. Yet it's one of the best ways to avoid getting sick. This simple habit requires only soap and warm water or an alcohol-based hand sanitizer — a cleanser that doesn't require water. Do you know the benefits of good hand hygiene and when and how to wash your hands properly?

6 6 Hand Washing Event (2/2) Proper hand-washing techniques Good hand-washing techniques include washing your hands with soap and water or using an alcohol-based hand sanitizer. Antimicrobial wipes or towelettes are just as effective as soap and water in cleaning your hands but aren't as good as alcohol-based sanitizers. Antibacterial soaps have become increasingly popular in recent years. However, these soaps are no more effective at killing germs than are regular soap and water. Using these soaps may lead to the development of bacteria that are resistant to the products' antimicrobial agents — making it even harder to kill these germs in the future. In general, regular soap is fine. The combination of scrubbing your hands with soap — antibacterial or not — and rinsing them with water loosens and removes bacteria from your hands. Proper hand washing with soap and water Follow these instructions for washing with soap and water: Wet your hands with warm, running water and apply liquid or clean bar soap. Lather well. Rub your hands vigorously together for at least 15 seconds. Scrub all surfaces, including the backs of your hands, wrists, between your fingers and under your fingernails. Rinse well. Dry your hands with a clean or disposable towel. Use a towel to turn off the faucet. Proper use of an alcohol-based hand sanitizer Alcohol-based hand sanitizers — which don't require water — are an excellent alternative to hand washing, particularly when soap and water aren't available. They're actually more effective than soap and water in killing bacteria and viruses that cause disease. Commercially prepared hand sanitizers contain ingredients that help prevent skin dryness. Using these products can result in less skin dryness and irritation than hand washing. Not all hand sanitizers are created equal, though. Some "waterless" hand sanitizers don't contain alcohol. Use only the alcohol-based products. To use an alcohol-based hand sanitizer: Apply about 1/2 tsp of the product to the palm of your hand. Rub your hands together, covering all surfaces of your hands, until they're dry. If your hands are visibly dirty, however, wash with soap and water rather than a sanitizer.

7 What‘s Hand Washing? How Mayo Clinic Guidelines Shape Upcoming Challenges for Machine Learning Models of Event Extraction Udo Hahn Jena University Language and Information Engineering (JULIE) Lab ECML PKDD 2008 European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases Workshop on High-Level Information Extraction September, Antwerp, Belgium

8 8 Approaches to Event Modeling Events as entities Events as flat relations –Features for flat events Decomposition into subevents Interleaving of subevents –Hard-wired ordinal connectivity –Triggered connectivity (e.g., via integrity constraints, inference rules, etc.) –Scripting for connectivity

9 9 Events as Entities unary relations... HandWashing Issues: –just ‚naming‘ of a relation (variant of ER) –no interrelations between any arguments

10 10 Flat Events Yet another breed of n-ary relations... HandWashing( Agent, Patient, Instruments, InState, OutState,...) Issues: –Are there ‚core‘ arguments (complement/adjunct)? –How many arguments (it‘s endless!)? –Type checking/compatibility

11 A remark on type constraints...

12 12 Features for Flat Events Issues: –Classifies verbs,..., use for inferences? –Classifies knowledge states (ongoing, result) – [the] splicing, [I am] splicing, [I] spliced, … HandWashing( Agent, Patient, Instruments, InState, OutState,...) Telicity: +/- (is there a point of completion ?) Aspectuality: n („I swim“ vs. „I am swimming“) Tense: n (location on time axis: past, now, future, …)

13 13 Hand Washing as a Complex Activity

14 14 Decomposition into Subevents HandWashing(...) Wet-w-Water(... ) ApplySoap(... ) RubHands(...) Rinse(...) Dry-w-Towel(...) Issues: –Abstraction between event cover and subevents –Subevent granularity –Subevent reusability –Completeness required ? Or are there mandatory vs. optional subevents ? Or are probabilities associated with subevents ? –Relevant vs. irrelevant intermediate steps the latter are often skipped in event descriptions

15 15 Interleaving of Subevents Hard-wired Ordinal Connectivity HandWashing(...) 1. Wet-w-Water(... ) 2. ApplySoap(... ) 3. RubHands(...) 4. Rinse(...) 5. Dry-w-Towel(...) Issues: –Orderings everywhere? Strict vs. partial Linear vs. parallel –Many orderings?: Defaults vs. exceptions

16 16 Interleaving of Subevents Triggered Connectivity (ICs, Rules) HandWashing(...) Wet-w-Water(... ) ApplySoap(... ) RubHands(...) Rinse(...) Dry-w-Towel(...) Issues: –Formal reasoning required IC checker Inference engine –How many relevant ICs/rules are there? A : hands fully soaped B : no more soap leftC : hands clean & dry A B C

17 17 Interleaving of Subevents Scripting for (Inter)Connectivity Hand Washing –... –Drying your hands Pre: hands are wet & no soap left Act: fetch towel –If towel not available call towel maintenance unit –Towel alternatives: other paper or textile tissues such as handkerchiefs, toilet paper,... Post: hands are dry & clean NonDefaultPre: hands are wet & alcohol-based hand sanitizer was used NonDefaultAct: wait until alcohol has evaporated Post: hand are dry & clean –If not clean: wash hands with soap Issues: –Massive knowledge acquisition bottleneck –Representation format –Doable at all ?

18 18 Lexical Encoding of Events Verbs (mind your [DG] parser!) –He washed his hands –I‘m washing my hands Perfective, progressive, aspectuality... but cf. also stative verbs such as know, like, … Nominalizations –My hand washing was a nightmare –Washing machines help you save time Adjectives, adverbs –Hand-washed shirts are cleaner than those which are machine-washed

19 19 Textual Encoding of Events Event cohesion –He washed his hands. They were covered with mud and thus needed extensive brushing. Event coherence –He washed his hands. The soap smelt like peaches. –He had washed his hands. Still the oil remained on his skin.

20 20 What‘s so Special about Events ? Moving from single, mostly binary relations to sets of interrelated n-ary relations (n usually >2) Types of interrelations: –Precedence/succession –(symbolic) temporal relations Temporal Interval Calculus: 13 atomic rels vs. time point (numerical clock) calculi –logical entailment   causality Event granularity Default events & (lots of) exceptions

21 21 Description of Static Knowledge –Logic (FOL, in particular) Description of Dynamic Knowledge (state changes) –Differential equations –Qualitative physics, biology, economics,... –Petri Nets (and other graph/network reps) –Planning languages (STRIPS, PDDL, ASBRU, …) –Logics considered harmful dynamic PL, TLs (point/interval), MLs Formal Description of Knowledge Types

22 22 (Some of) The Challenges of Event Representation Associating ontologies w.\ textual realizations –Linguistic categories (on-going process, result of a process, etc.) matter  features Scalability from simple to advanced representations (granularity sliding) Different breed of inference rules –real ‚world modeling‘ Frame axioms (tracking changes of the world) Given all that representational richness – How tractable are event calculi? Stay on the poor side of life ? – How adequate will your results be ?

23 23 Machine Learning Challenges of Event Extraction Formal basis of event description –Symbolic, discrete KRs Learning the building blocks of complex events –Sets of n-ary relations Learning connectivity criteria for these (sub)relations –Precedence/succession –Temporal orderings –ICs, Inference rules   “Causality” –Numerical, continuous KRs Quantitative data   induction of differential equations Methodological Frameworks –Learning (timed, probabilistic) FSAs, Bayesian Ns –Event (process) mining: Petri Nets –ILP –Temporal logic-based learning (see also TimeML)

24 24 Clinical Events: Guidelines National Institute for Health and Clinical Excellence (for diabetis – type 2)

25 25 Clinical Events: Formalization of Guidelines Y. Shavar, S. Miksch, P. Johnson AI in Medicine, 1997 G. Duftschmidt, S. Miksch, W. Gall AI in Medicine, 2002

26 26 Biological Events: Gene Regulation Figure: Positive and Negative Regulation of Gene Transcription (Expression). The process that modulates the frequency, rate or extent of gene expression, where gene expression is the process in which the coding sequence of a gene is converted into a gene product(s) (protein, RNA). (Gene Ontology)‏

27 What‘s an Event ? How Ontologies and Linguistic Semantics Shape Upcoming Challenges for Machine Learning Models of Event Extraction Udo Hahn Jena University Language and Information Engineering (JULIE) Lab ECML PKDD 2008 European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases Workshop on High-Level Information Extraction September, Antwerp, Belgium


Download ppt "What‘s an Event ? How Ontologies and Linguistic Semantics Shape Upcoming Challenges for Machine Learning Models of Event Extraction Udo Hahn Jena University."

Similar presentations


Ads by Google