1 Information pragmatics Lessons for the web from human languages Christopher Manning Stanford DB seminar September 2000 http://nlp.stanford.edu/~manning/
Dec 19, 2015
1
Information pragmaticsLessons for the web from human
languages
Christopher ManningStanford DB seminar
September 2000http://nlp.stanford.edu/~manning/
2
The problem
• When people see web pages, they understand their meaning – By and large. To the extent that they don’t,
there’s a gradual degradation• When computers see web pages, they get only
character strings and HTML tags
3
The human view
4
The intelligent agent view
<HTML> <HEAD><TITLE>Ford Motor Company - Home Page</title><META NAME="Keywords" CONTENT="cars, automobiles, trucks, SUV,
mazda, volvo, lincoln, mercury, jaguar, aston martin, ford"><META NAME="description" CONTENT="Ford Motor Company corporate
home page"><SCRIPT LANGUAGE="JavaScript1.2"> … </SCRIPT><!-- Trustmark code --><DIV ID=trustmarkDiv> <TABLE BORDER="0" CELLPADDING=0 CELLSPACING=0 WIDTH=768> <TR><TD WIDTH=768 ALIGN=CENTER> <A HREF="default.asp?
pageid=473" onmouseover="logoOver('fordscript');rolloverText('ht0')" onmouseout="logoOut('fordscript');rolloverText('ht0')"><img border="0" src="images/homepage/fordscript.gif" ALT="Learn more about Ford Motor Company" WIDTH="521" HEIGHT="39"></A><br>
… </TD></TR></TABLE></DIV> </BODY></HTML>
5
The problem (cont.)
• We'd like computers to see meanings as well, so that computer agents could more intelligently process the web
• These desires have led to XML, RDF, agent markup languages, and a host of other proposals and technologies which attempt to impose more syntax and semantics on the web – in order to make life easier for agents.
6
Ontologies
The answer, it is suggested, is ontologies• Shared formal conceptualizations of particular
domains [concepts, relations, objects, and constraints]
• An ontology is a specification of a conceptualization that is designed for reuse across multiple applications
• Ontologies: controlled vocabularies, taxonomy, OO database schema, knowledge-representation system
• Ontologies, as specifications of the concepts in a given field, and of the relationships among those concepts, provide insight into the nature of information produced by that field and are an essential ingredient for any attempts to arrived at a shared understanding of concepts in a field.
7
Why is this idea appealing?
• An ontology is really a dictionary. A data dictionary.
• In the world of closed company databases, one had a clear semantics for fields and tables, and the ability to combine information across them by well-specified logical means
• In the world-wide web, you have a mess• The desire for a global or industry-wide
ontology is a desire to bring back the good old days.
8
Thesis
• The problem can’t and won’t be solved by mandating a universal semantics for the web.
9
Nuanced Thesis (1)
• Structured knowledge is important, and there will be increasing use of structure and keys … just as we started using zipcodes, and then the postoffice started barcoding.
• These processes all offer the opportunity to increase speed and precision, and agents will want to use them when available and reliable
• But successful agents will need to be able to work even when this information isn’t there.
• The postoffice still delivers your mail, even when the zipcode is missing … or wrong.
10
Nuanced Thesis/Theses? (2)
• There will never be a complete explicit and unambiguous semantics for everything needed on the web … or even a non-trivial chunk of it … both because of the scale of the problem and the speed of change
• Much of the semantic knowledge needs instead to reside in the agent
• The agent needs to be able to reason with imperfect and inexact information found in the world (or on the web, if you will) using context as well as its own knowledge
11
XML?
• I’m not saying that XML won’t be used much. It certainly will be used widely
• Internally, it will be used for most content (except tabular data), so that content can be easily retargeted for browsers, WAP, iMode, and whatever comes next
• Some sites will publish XML to outside users.• But a lot won’t.
12
Will XML be published?
• “Another lesson of transitions is that the old way persists for a very long time. The 4.0-level browsers will be with us for the foreseeable future.” – Dave Winer (reacting to similar conclusions of Jakob Nielsen)
• If you’re going to be serving HTML for “the foreseeable future” [2003?], why bother complicating your life by serving something else as well?
• Especially when it doesn’t look better to the user
13
XML
• Even when it is published, XML goes only a small way to enabling knowledge transfer
• It is simply a syntax• The same meanings can be encoded by it in
many ways, and conversely, different meanings can be coded in the same way.
• This is what suggests the need for a clearly mandated semantics for web markup
14
Explicit, usable web semantics
• Will such a thing work?• That is, will web pages be consistently marked
up with a uniform explicit semantics that is easily processed by agents so that they don’t have to deal with that messy HTML that underlies what humans look at?
• I think not.• For a bunch of reasons.
15
(1) The semantics
• Are there adequate and adequately understood methods for marking up pages with such a consistent semantics, in such a way that it would support simple reasoning by agents?
• No.
16
What are some AI people saying?
“Anyone familiar with AI must realize that the study of knowledge representation—at least as it applies to the “commensense” knowledge required for reading typical texts such as newspapers—is not going anywhere fast. This subfield of AI has become notorious for the production of countless non-monotonic logics and almost as many logics of knowledge and belief, and none of the work shows any obvious application to actual knowledge-representation problems. Indeed, the only person who has had the courage to actually try to create large knowledge bases full of commonsense knowledge, Doug Lenat …, is believed by everyone save himself to be failing in his attempt.” (Charniak 1993:xvii–xviii)
17
(2) Many of the problems are pragmatics not semantics
pragmatic relating to matters of fact or practical affairs often to the exclusion of intellectual or artistic matters
pragmatics linguistics concerned with the relationship of the meaning of sentences to their meaning in the environment in which they occur
• A lot of the meaning in web pages (as in any communication) derives from the context – what is referred to in the philosophy of language tradition as pragmatics
• Communication is situated
18
The crêperie
• After making use of 3 different picture search engines, and spending at least ½ an hour on the site of a very dedicated French photographer, I had found the setting for my story … a crêperie.
• Well, almost. The visuals didn’t really convey what I needed, so let me settle for a worse quality picture of a gyro shop.
19
Not actually a crêperie
20
Important points
• “Multimedia” information sources are vital• The meaning of a ‘text’ is strongly determined
by its context of use• Indeed, you can think of language as conveying
the minimal amount of information necessary given the context and assumed shared knowledge
• Humans are used to communicating even when they don’t completely hear or understand the signal [even if this example is a bit extreme]
21
Pragmatics on the web
• Information supplied is incomplete – humans will interpret it– Numbers are often missing units– A “rubber band” for sale at a stationery site is a
very different item to a rubber band on a metal lathe
– A “sidelight” means something different to a glazier than to a regular person
• Humans will evaluate content using information about the site, and the style of writing– value filtering
22
(3) The world changes
• The way in which business is being done is changing at an astounding rate– or at least that’s what the ads from e
business companies scream at us• Semantic needs and usages evolve (like
languages) more rapidly than standards (cf. the Académie française)
People use words that aren’t in the dictionary.Their listeners understand them.
23
Rapid change
• Last year Rambus wasn’t a concept in computer memory classification, now it is
• Cell phones have long had attributes like size and battery life– Now whether they support WAP is an attribute– In a couple of years time that attribute will
probably have disappeared again
People will introduce new products when they’re ready, not when some committee has added the terms to an ontology
24
(4) Interoperation
Ontology: a shared formal conceptualization of a particular domain
• Meaning transfer frequently has to occur across the subcommunities that are currently designing *ML languages, and then all the problems reappear, and the current proposals don't do much to help
25
Many products cross industries
http://www.interfilm-usa.com/Polyester.htm
• Interfilm offers a complete range of SKC's Skyrol® brand polyester films for use in a wide variety of packaging and industrial processes.
• Gauges: 48 - 1400• Typical End Uses: Packaging, Electrical, Labels,
Graphic Arts, Coating and Laminating– labels: milk jugs, beer/wine, combination
forms, laminated coupons, …
26
Mismatches
• When interoperation involves distinct domains or just distinct subcommunities within an industry, semantic mismatch ensues
• Local representational power conflicts with global consistency [you want to advertise your new feature]– Your own needs will take priority
• Systems will need to deal with this heterogeneity• Integration of information across XML markup
languages is scarcely easier than integration of the same information represented in HTML.
27
Semantic mismatches
Different Usages• Cell phone = mobile phone• Data projector = beamerDifferent levels of specialized vocabulary• “water table” = the strip of wood that points
outward at the bottom of the door– [hydrologists mean something very different by “water
table”]
Ambiguity of reference• Is “C.D. Manning” the same person as
“Christopher Manning”?
28
Name matching/Object identity knowledge
• Database theory is built around ideas of unique identifiers, determinate relational operations, …
• (Human) natural language processing is built around context-embedded reasoning about issues of identity and meaning– In the Stanford DB group, “Hector” means …– In a course on Homer, “Hector” means ...
• Integrating information sources requires probabilistic reasoning about object identity
29
Text learning: A little experiment
• Take a large quantity of Wall Street Journal newswire– articles are tagged for company ticker
symbols• Gather articles for each ticker symbol• Do proper name delimiting• Collect distinctively common proper names for
subclass– so all companies aren’t called “Wall Street
Journal”
30
Results
• IBM• International Business
Machines Corp.• International Business
Machines
• Weyerhaeuser Co.• Weyerhaeuser• Weyerhaeuser Canada
Ltd.• Weyerhaeuser Canada• Tacoma
• Alcatel• Alcatel-Alsthom SA• Alcatel-Alsthom• Alcatel-Alsthom NV
• General Electric• General Electric Co.• GE• GE Capital Corp.• GE Capital• General Electric Capital
Corp.
31
Limitations in accuracy
• All such techniques are going to be imperfect and produce some errors
• But:– life is imperfect; there are always errors and
missing data– you do get confidence factors which among
other things can be used to tradeoff accuracy
32
Recall vs. Precision
• High recall:– You get all the right answers, but garbage too.– Good when incorrect results are not problematic.– Common from automatic systems.
• High precision:– When all returned answers must be correct.– Good when missing results are not problematic.– Common from hand-built systems.
• In general in these things, one can trade one for the other– But it’s hard to score well on both
precision
recall
x
x
x
x
33
(5) Pain but no gain
• A lot of the time people won’t put in information according to standards for semantic/agent markup, even if they exist.
• Three reasons…
34
(5.1) Pain no gain
Laziness: • Only 0.3% of sites currently use the (simple)
Dublin Core metadata standard. (Lawrence and Giles 1999).
• Even less are likely to use something that is more work
• Why? They don’t appear to perceive much value I guess. What would change this?
35
Inconsistency: digital cameras
• Image Capture Device: 1.68 million pixel 1/2-inch CCD sensor• Image Capture Device Total Pixels Approx. 3.34 million
Effective Pixels Approx. 3.24 million• Image sensor Total Pixels: Approx. 2.11 million-pixel• Imaging sensor Total Pixels: Approx. 2.11 million 1,688 (H) x
1,248 (V)• CCD Total Pixels: Approx. 3,340,000 (2,140[H] x 1,560 [V] )
– Effective Pixels: Approx. 3,240,000 (2,088 [H] x 1,550 [V] )– Recording Pixels: Approx. 3,145,000 (2,048 [H] x 1,536 [V] )
• These all came off the same manufacturer’s website!!
• And this is a very technical domain. Try sofa beds.
36
(5.2) Pain no gain
• “Sell the sizzle, not the steak”• The way businesses make money is by selling
something at a profit (for more than necessary)
• The way you do this is by getting people to want it from you:– advertising– site stickiness (“while I’m here…”)– trust
Newspaper advertisements rarely contain spec sheets
37
(5.2) Pain no gain
• Having an easily robot-crawlable site is a recipe for turning what you sell into a commodity
• This may open new markets• But most would prefer not to be in this
business • Having all your goods turned into a commodity
by a shopping bot isn’t in your best interest.– the profits are very low
38
(5.3) Gain, no pain
• The web is a nasty free-wheeling place• There are people out there that will abuse the
intended use and semantics of any standard, providing they see opportunities to profit from doing so
• An agent cannot simply believe the semantics• It will have to reason skeptically based on all
contextual and world knowledge available to it.
39
(6) Less structure to come
• “the convergence of voice and data is creating the next key interface between people and their technology. By 2003, an estimated $450 billion worth of e-commerce transactions will be voice-commanded.*”
• Question: will these customers speak XML tags?
Intel ad, NYT, 28 Sep 2000*Data Source: Forrester Research.
40
Summary so far
• With large-scale distributed networked information sources like the web, everyone suddenly needs to deal with highly heterogeneous data sources of uncertain correctness and value, where there are frequent semantic mismatches in which terms are used or what they mean. Contextual information is often needed to determine the meaning or reference of terms. In other words, the problems look a lot like Natural Language Processing, regardless of whether the data is text as narrowly defined.
41
The connection to language
Decker et al. IEEE Internet Computing (2000):
• “The Web is the first widely exploited many-to-many data-interchange medium, and it poses new requirements for any exchange format:– Universal expressive power– Syntactic interoperability– Semantic interoperability”
But human languages have all these properties, and maintain superior expressivity and interoperability through their flexibility and context dependence
42
The direction to go
• Successful agents will need prior knowledge, and use ontologies, etc. to help interpret web pages – they become a locus of semantics
• But they will also depend on contextual knowledge and reasoning in the face of uncertain information.
• They will use well-marked up information, if available and trusted, but they will be able to extract their own metadata from information intended for humans, regardless of the form in which the information appears.
43
The scale of the problem
• The web is too big a thing for it to be likely for humans to hand-enter metadata for most pages
• Hand-building ontologies and reasoning systems hasn’t been very successful
• Agents must be able to extract propositions or relations from information intended for humans
• A useful observation in seeking this goal is that text statistics can often be used as a surrogate for world knowledge
44
Processing textual data
Use language technology to add value to data by:• interpretation• transformation• value filtering• augmentation (providing metadata)
Two motivations:• The large amount of information in textual form• Information integration needs NLP-style methods
45
Knowledge Extraction VisionMulti-
dimensional Meta-data Extraction
J F M A M J J A
EMPLOYEE / EMPLOYER Relationships:Jan Clesius works for Clesius EnterprisesBill Young works for InterMedia Inc.COMPANY / LOCATION Relationshis:Clesius Enterprises is in New York, NYInterMedia Inc. is in Boston, MA
Meta-Data
India Bombing NY Times Andhra Bhoomi Dinamani Dainik Jagran
Topic Discovery
Concept Indexing
Thread Creation
Term Translation
Document Translation
Story Segmentation
Entity Extraction
Fact Extraction
46
Task: Text Categorization
• Take a document and assign it a label representing its content.
• Classic example: decide if a newspaper article is about politics, business, or sports?
• But there are many relevant web uses for the same technology:– Is this page a laser printer product page?– Does this company accept overseas orders?– What kind of job does this job posting describe?– What kind of position does this list of responsibilities
describe?– What position does this this list of skills best fit?
47
Text Categorization
• Usually, simple machine learning algorithms are used.• Examples: Naïve Bayes models, decision trees.• Very robust, very re-usable, very fast.• Recently, slightly better performance from better
algorithms– e.g., use of boosting, support vector machines
• Accuracy is more dependent on:– Naturalness of classes.– Quality of features extracted and amount of training
data available.• Accuracy typically ranges from 65% to 95% depending
on the situation.
48
Task: Information Extraction
Suppositions:• A lot of information that could be represented
in a structured semantically clear format isn’t• It may be costly, not desired, or not in one’s
control (screen scraping) to change this.
• Goal: being able to answer semantic queries using “unstructured” natural language sources
49
Information Extraction
• Information extraction systems– Find and understand relevant parts of texts.– Produce a structured representation of the relevant
information: relations (in the DB sense)• These systems:
– Combine knowledge about language and the application domain
– Automatically extract the desired information• When is IE appropriate?
– Clear, factual information (who did what to whom and when?)
– Only a small portion of the text is relevant.
50
Task: Wrapper Induction
• IE– Linguistic features let us figure out relations in the
data.– Relations expressed by the text.
• Wrapper Induction– Sometimes, the relations are structural.
• Webpages generated by a database.• Tables, lists, etc.
– Relations can be expressed by the structure of the document:
• the item in bold in the 3rd column of the table is the price
51
Wrapper Induction
• Handcoding a wrapper in Perl isn’t very viable– sites are numerous, and their surface structure
mutates rapidly
• One wants to learn rules using machine learning methods
• Wrapper induction techniques can learn not only about tabular data but things like:– If there is a page about a research project X and
there is a link near the word ‘people’ to a page that is about a person Y then Y is a member of the project X.
– [e.g, Tom Mitchell’s Web->KB project]
52
Example: Classified Advertisements (Real Estate)Background:• Advertisements are plain text• Lowest common denominator: only thing that
70+ newspapers with 20+ publishing systems can all handle
Goals:• Extraction of information for input into
database tables• To permit searching and other functionality on
a web site
53
Unstructured Text Input
<ADNUM>2067206v1</ADNUM><DATE>March 02, 1998</DATE><ADTITLE>MADDINGTON $89,000</ADTITLE><ADTEXT>OPEN 1.00 - 1.45<BR>
U 11 / 10 BERTRAM ST<BR> NEW TO MARKET Beautiful<BR> 3 brm freestanding<BR> villa, close to shops & bus<BR> Owner moved to Melbourne<BR> ideally suit 1st home buyer,<BR> investor & 55 and over.<BR> Brian Hazelden 0418 958 996<BR> R WHITE LEEMING 9332 3477
</ADTEXT>
54
Real Estate Ads: Output
• Output is database tables• But the general idea in slot-filler format:
SUBURB: MADDINGTONADDRESS: (11,10,BERTRAM,ST)INSPECTION: (1.00,1.45,11/Nov/98)BEDROOMS: 3TYPE: HOUSEAGENT: BRIAN HAZELDENBUS PHONE: 9332 3477MOB PHONE: 0418 958 996
• Sometimes need to expand “quantification”
55
57
Why doesn’t text search (IR) work?
Consider what you want to search for in real estate advertisements:
• Suburbs. You might think easy, but:– Mentioned in other contexts
• Coldwell Banker, Mosman• Only 45 minutes from Parramatta• Rooty Hill, the Paddington of the West
– Multiple property ads have different suburbs
58
Why doesn’t text search (IR) work?
• Money– People are looking for a monetary range, not
a textual match– Multiple money amounts
• was $155,000, now $145,000– Various forms:
• offers in the high 70s [not rents for $120]• Bedrooms
– Similar issues (range, various forms)• br, bdr, b/room, b.r., brm, B/R, beds, …
59
Text categorization
• Many ads in listings aren’t correctly classified:– e.g., Houses for sale section includes ads
from people who value or inspect houses– Small paper may not differentiate
categories (home/business for sale)
60
Complex tokens and chunk parsing of constituents
• Parses complex tokens and sentence chunks:– Noun groups, prepositional phrases, verb-
complement patterns, addresses, …– Finite state constituency
• This is sufficient to filter out most false cues: n– was $70,000
• and to parse up complex expressions:– “in the high sixties” = $67,000 - $69,999
• and special forms, such as:– addresses, inspection details, etc.
61
Semantic Coding
• Monetary amounts, percentages, etc. are easy• Name recognition is pretty easy (capital letters)
– But there are towns: Sale, Alice, Price, Home Hill
• Semantic type is harder (company, place, …)– Often use external knowledge sources
• Hard thing is abbreviations, acronyms, variants– the object identity problem again!– Alexandra Heights vs. Alex Heights– Kangaroo Point vs. K’ Pt– GREENVALLEY ?= GREEN VALLEY
62
Text Segmentation
A single ad will often advertise multiple propertiesSOUTHPORT UNIT SPECIALS$58,900 o.n.o. 2 brm close to water and shops.$114,000 "Grandview", excellent value, good returnsLJ Coleman Real EstateContact Steve 5527 0572
GLEBE 2br yd $250; 4br yd $430
COOGEE 3br yd $320; 1br $150
BALMAIN 1br $180
H.R. Licensed FEE 9516-3211
63
Real estate ad extraction
• The system works reliably and accurately (over 90% accuracy on each element (suburb, price …)– segmentation errors still cause trouble: a
mistake means a lot of information is wrong• Well-suited to a web application such as this:
– Rapid development and deployment– Gets vast bulk of ads into the system quickly – There are some errors, but since web service is
free addition, it doesn’t have to be perfect
64
Precision & Semantic markup
The story so far:• We can get a fair way with text learning!• In some places, moderate accuracy is okay
• But often business needs precision – as Gio Wiederhold points out in his talks
• These methods may not offer sufficient accuracy
65
Precision & Semantic markup
• This is where semantic markup comes back in• If a page has reliable semantic markup, such a
program can use it to provide much higher accuracy levels
• Agents will need to check the provided markup• But deciding that provided semantic markup is
trustworthy is a lot easier (and hence more reliable) decision than working out the meaning from unstructured text
66
Data verification
• Humans are very good at checking if data is reasonable:– 5525 Beverly Place,
Pittsburgh– 361-5525
• They know if content is reasonable by content analysis
67
Data verification
• Most programs are dumb– especially if they expect to just rely on
semantic markup• Again one needs unstructured text
classification and learning– one needs to check that field contents are
reasonable• Richly semantically marked up data has a real
use here, since it allows agents to continue to learn (especially as usage changes over time)
68
Conclusion
• Rich semantic markup has an important place: improving the precision of agent understanding
• But there will be no substitute for agents that can work with “unstructured” data– part of that data is text [what I know about!]– but visual and other information is also
incredibly important• one really needs to use how a page looks
• All of it involves reasoning from uncertain situated information more in the style of NLP
69