Posted on

Background for our Semantic Technology 2013 presentation

In the spring of 2012, Vulcan engaged Automata for a knowledge acquisition (KA) experiment.  This article provides background on the context of that experiment and what the results portend for artificial intelligence applications, especially in the areas of education.  Vulcan presented some of the award-winning work referenced here at an AI conference, including a demonstration of the electronic textbook discussed below.  There is a video of that presentation here.  The introductory remarks are interesting but not pertinent to this article.

Background on Vulcan’s Project Halo

Background on Vulcan's Project Halo

From 2002 to 2004, Vulcan developed a Halo Pilot that could correctly answer between 30% and 50% of the questions on advanced placement (AP) tests in chemistry.  The approaches relied on sophisticated approaches to formal knowledge representation and expert knowledge engineering.  Of three teams, Cycorp fared the worst and SRI fared the best in this competition.  SRI’s system performed at the level of scoring a 3 on the AP, which corresponds to earning course credit at many universities.  The consensus view at that time was that achieving a score of 4 on the AP was feasible with limited additional effort.  However, the cost per page for this level of performance was roughly $10,000, which needed to be reduced significantly before Vulcan’s objective of a Digital Aristotle could be considered viable.

Continue reading “Background for our Semantic Technology 2013 presentation”

Posted on

Semantic Technology & Business Conference (SemTechBiz)

Benjamin Grosof and I will be presenting the following review of recent work at Vulcan towards Digital Aristotle as part of Project Halo at SemTechBiz in San Francisco the first week of June.

Acquiring deep knowledge from text

We show how users can rapidly specify large bodies of deep logical knowledge starting from practically unconstrained natural language text.

English sentences are semi-automatically interpreted into  predicate calculus formulas, and logic programs in SILK, an expressive knowledge representation (KR) and reasoning system which tolerates practically inevitable logical inconsistencies arising in large knowledge bases acquired from and maintained by distributed users possessing varying linguistic and semantic skill sets who collaboratively disambiguate grammar, logical quantification and scope, co-references, and word senses.

The resulting logic is generated as Rulelog, a draft standard under W3C Rule Interchange Format’s Framework for Logical Dialects, and relies on SILK’s support for FOL-like formulas, polynomial-time inference, and exceptions to answer questions such as those found in advanced placement exams.

We present a case study in understanding cell biology based on a first-year college level textbook.

Posted on

Project Sherlock

Working as part of Vulcan’s Project Halo[1], Automata is applying a natural language understanding system that translates carefully formulated sentences into formal logic so as to answer questions that typically require deeper knowledge and inference than demonstrated by Watson.

The objective over the next three quarters is to acquire enough knowledge from the 9th edition of Campbell’s Biology textbook to demonstrate three things.

  • First, that the resulting system answers, for example, biology advanced placement (AP) exam questions more competently than existing systems (e.g., Aura[2] or Inquire[3]).
  • Second, that knowledge from certain parts of the textbook is effectively translated from English into formal knowledge with sufficient breadth and depth of coverage and semantics.
  • Third, that the knowledge acquisition process proves efficacious and accessible to less than highly skilled knowledge engineers so as to accelerate knowledge acquisition beyond 2012.

Included in the second of these is a substantial ontology of background knowledge expected of students in order to comprehend the selected parts of the textbook using a combination of OWL, logic, and English sentences from sources other than the textbook.

Automata is hiring logicians, linguists, and biologists to work as consultants, contracts, or employees for:

  • Interactive tree-banking and word-sense disambiguation of several thousand sentences.[4]
  • Extending its lexical ontology and a broad-coverage grammar of English with additional vocabulary and deeper semantics, especially concerning cellular biology and related scientific knowledge including chemistry, physics, and math.
  • Maturing its upper and middle ontology of domain independent knowledge using OWL in combination with various other technologies, including description logic, first-order logic, high-order logic, modal logic, and defeasible logic.[5]
  • Enhancing its platform for text-driven knowledge engineering towards a collaborative wiki-like architecture for self-aware content in scientific education and biomedical applications.

Terms of engagement are flexible; ranging from small units of work to full-time employment.  We are based in Pittsburgh, Pennsylvania and Vulcan is headquartered in Seattle, Washington, but the team is distributed across the country and overseas.

Please contact Paul Haley by e-mail to his first name at this domain.

[1] Vulcan: http://www.vulcan.com/TemplateCompany.aspx?contentId=54; Project Halo: http://www.projecthalo.com/
Video introduction/overview :http://videolectures.net/aaai2011_gunning_halobook/

[2] Aura: http://www.ai.sri.com/project/aura

[3] Inquire: http://www.franz.com/success/customer_apps/artificial_intelligence/aura.lhtml

[4] tree-banking and WSD: http://www.omg.org/spec/SBVR and http://en.wikipedia.org/wiki/Word-sense_disambiguation

[5] e.g., SILK (http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.174.1796 ) and SBVR (http://www.omg.org/spec/SBVR)

Posted on

NLP: depictive in an HPSG lexicon?

We’re working with the English Resource Grammar (ERG), OWL, and Vulcan’s SILK to educate the machine by translating textbooks into defeasible logic.  Part of this involves an ontology that models semantics more deeply than the ERG, which is based on head-driven phrase structure grammar (HPSG), which provides deeper parsing and, with the ERG and the DELPH-IN infrastructure, also provides a simple under-specified semantic representation called minimal recursion semantics (MRS).

We’re having a great time using OWL to clarify and enrich the semantics of the rich model underlying the ERG.  Here’s an example, FYI.  If you’d like to know more (or help), please drop us a line!  Overall the project will demonstrate our capabilities for transforming everyday sentences into RIF and business rule languages using SBVR extended with defeasibility and other capabilities, all modeled in the same OWL ontology.

What triggered this blog entry was a bit of a surprise in seeing that whether or not an adjective could be used depictively is sometimes encoded in the lexicon.  This is one of the problems of TDL versus a description-logic based model with more expressiveness.  It results in more lexical entries than necessary, which has been discussed by others when contrasted with the attributed logic engine (ALE), for example.

In trying to model the semantics of words like ‘same’ and ‘different’, we are scratching our heads about these lines from the ERG’s lexicon:

  1. same_a1 := aj_pp_i-cmp-sme_le & [ ORTH < “same” >, SYNSEM [ LKEYS.KEYREL.PRED “_same_a_as_rel”, …
  2. the_same_a1 := aj_-_i-prd-ndpt_le & [ ORTH < “the”, “same” >, SYNSEM [ LKEYS.KEYREL.PRED “_the+same_a_1_rel”, …
  3. the_same_adv1 := av_-_i-vp-po_le & [ ORTH < “the”, “same” >, SYNSEM [ LKEYS.KEYREL.PRED “_the+same_a_1_rel”, …
  4. exact_a2 := aj_pp_i-cmp-sme_le & [ ORTH < “exact” >, SYNSEM [ LKEYS.KEYREL.PRED “_exact_a_same-as_rel”…

One of the interesting things about lexicalized grammars is that lexical entries (i.e., ‘words’) are described with almost arbitrary combinations of their lexical, syntactic, and semantic characteristics.

The preceding code is expressed in a type description language (TDL) used by the Lisp-based LKB (and its C++ counterpart, PET, which are unification-based parsers that produce a chart of plausible parses with some efficiency.  What is given above is already deeper than what you can expect from a statistical parser (but richer descriptions of lexical entries promises to make statistical parsing much better, too).

Unfortunately, there is no available documentation on why the ERG was designed as it is, so the meaning of the above is difficult to interpret.  For example, the types of lexical entries (the symbols ending in ‘_le’) referenced above are defined as follows:

  1. aj_pp_i-cmp-sme_le := basic_adj_comp_lexent & [SYNSEM[LOCAL[CAT[HEAD superl_adj &[PRD -,MOD <[LOCAL.CAT.VAL.SPR <[–MIN def_or_demon_q_rel]>]>],VAL.SPR.FIRST.–MIN much_deg_rel],CONT.RELS <!relation,relation!>],MODIFD.LPERIPH bool,LKEYS[ALTKEYREL.PRED comp_equal_rel,–COMPKEY _as_p_comp_rel]]].
  2. aj_-_i-prd-ndpt_le := nonc-hm-nab & [SYNSEM basic_adj_abstr_lex_synsem & [LOCAL[CAT[HEAD adj & [PRD +,MINORS[MIN norm_adj_rel,NORM norm_rel],TAM #tam,MOD < anti_synsem_min >],VAL[SPR.FIRST anti_synsem_min,COMPS < >],POSTHD +],CONT[HOOK[LTOP #ltop,INDEX #arg0 &[E #tam],XARG #xarg],RELS <! #keyrel & adj_relation !>,HCONS <! !>]],NONLOC non-local_none,MODIFD notmod &[LPERIPH bool],LKEYS.KEYREL #keyrel &[LBL #ltop,ARG0 #arg0,ARG1 #xarg & non_expl-ind]]].

Needless to say, that’s a mouthful!  Chasing this down, the following ‘informs’ us that “the same”, which uses type #2 above, is defined using the following lexical types:

  1. nonc-hm-nab := nonc-h-nab & mcna.
  2. nonc-h-nab := nonconj & hc-to-phr & non_affix_bearing.
  3. mcna := word & [ SYNSEM.LOCAL.CAT.MC na ].

Which is to say that it is non-conjunctive, complements a head to form a phrase, can’t be affixed, cannot constitute a main clause, and is a word.

The fact that the lexical entry for “the same” is adjectival is given the definition of the following type(s) used in the SYNSEM feature:

  1. basic_adj_comp_lexent := compar_superl_adj_word & [SYNSEM adj_unsp_ind_twoarg_synsem & [LOCAL[CAT.VAL[COMPS <canonical_or_unexpressed & [–MIN #cmin,LOCAL [CAT basic_pp_cat,CONJ cnil,CONT.HOOK [LTOP #ltop,INDEX #ind]]]>],CONT.HOOK [ LTOP #ltop, XARG #xarg]],LKEYS [ KEYREL.ARG1 #xarg,ALTKEYREL.ARG2 #ind,–COMPKEY #cmin]]].b
  2. compar_superl_adj_word := nonc-hm-nab & [SYNSEM adj_unsp_ind_synsem & [LOCAL[CAT[HEAD[MOD <[–SIND #ind & non_expl]>,TAM #tam,MINORS.MIN abstr_adj_rel],VAL.SPR.FIRST.LOCAL.CONT.HOOK.XARG #altarg0],CONT[HOOK[XARG #ind,INDEX #arg0 & [E #tam]],RELS.LIST <[LBL #hand,ARG1 #ind],#altkeyrel & [LBL #hand,ARG0 event & #altarg0,ARG1 #arg0],…>]],LKEYS.ALTKEYREL #altkeyrel]].

Which is to say that it is a comparative or superlative adjectival word (even though it consists of two lexemes in its ‘orthography’) that involves two semantic arguments including one complement which may be unexpressed prepositional phrase.  A comparative or superlative adjective, in turn, is non-conjunctive, complements a head to form a phrase, is non-affix bearing (?), and non-clausal, as defined by the type ‘nonc-hm-nab’ above.

The types used in the syntax and semantic (i.e., SYNSEM) feature of the two lexical types are defined as follows (none of which is documented):

  1. adj_unsp_ind_twoarg_synsem := adj_unsp_ind_synsem & two_arg.
  2. adj_unsp_ind_synsem := basic_adj_lex_synsem & lex_synsem & adj_synsem_lex_or_phrase & isect_synsem & [LOCAL.CONT.HOOK.INDEX #ind,LKEYS.KEYREL.ARG0 #ind].

In a moment, we’ll discuss the types used in the second of these, but first, some basics on the semantics that are mixed with the syntax above.

In effect, the above indicates that a new ‘elementary predication’ will be needed in the MRS to represent the adjectival relationship in the logic derived in the course of parsing (i.e., that’s what ‘unsp_ind’ means, although it’s not documented, which I will try not to bemoan much further.)

The following indicates that the newly formed elementary predicate is not (initially) within any scope and that it has two arguments whose semantics (i.e., their RELations) are concatenated for propagation into the list of elementary predications that will constitute the MRS for any parses found.

  1. two_arg := basic_two_arg & [LOCAL.CONT.HCONS <! !>].
  2. basic_two_arg := unspec_two_arg & lex_synsem.
  3. unspec_two_arg := basic_lex_synsem & [LOCAL.ARG-S <[LOCAL.CONT.HOOK.–SLTOP #sltop,NONLOC [SLASH[LIST #smiddle,LAST #slast],REL [LIST #rmiddle,LAST #rlast],QUE[LIST #qmiddle,LAST #qlast]]],[LOCAL.CONT.HOOK.–SLTOP #sltop, NONLOC[SLASH[LIST #sfirst,LAST #smiddle],REL[LIST #rfirst,LAST #rmiddle],QUE[LIST #qfirst,LAST #qmiddle]]]>,LOCAL.CONT.HOOK.–SLTOP #sltop,NONLOC[SLASH[LIST #sfirst,LAST #slast],REL[LIST #rfirst,LAST #rlast],QUE[LIST #qfirst,LAST #qlast]]].
  4. lex_synsem := basic_lex_synsem & [LEX +].

The last of these expresses that the constuction is lexical rather than phrasal (which includes clausal in the ERG).

Continuing with the definition of “the same” as an adjective, the following finally clarifies what it means to be a basic adjective:

  1. basic_adj_lex_synsem := basic_adj_abstr_lex_synsem & [LOCAL[ARG-S <#spr . #comps>,CAT[HEAD adj_or_intadj,VAL[SPR<#spr & synsem_min &[–MIN degree_rel,LOCAL[CAT[VAL[SPR *olist*,SPEC <[LOCAL.CAT.HS-LEX #hslex]>],MC na],CONT.HOOK.LTOP #ltop],NONLOC.SLASH 0-dlist,OPT +],anti_synsem_min &[–MIN degree_rel]>,COMPS #comps],HS-LEX #hslex],CONT.RELS.LIST <#keyrel,…>],LKEYS.KEYREL #keyrel & [LBL #ltop]].

Well, ‘clarifies’ might not have been the right word!  Essentially, it indicates that the adjective may have an optional degree specifier (which semantically modifies the predicate of the adjective) and that the predicate specified in the lexical entry becomes the predicate used in the MRS.  The rest is defined below:

  1. basic_adj_abstr_lex_synsem := basic_adj_synsem_lex_or_phrase & abstr_lex_synsem & [LOCAL.CONT.RELS.LIST.FIRST basic_adj_relation].
  2. basic_adj_synsem_lex_or_phrase := canonical_synsem & [LOCAL[AGR #agr,CAT[HEAD[MINORS.MIN basic_adj_rel],VAL[SUBJ <>,SPCMPS <>]],CONT.HOOK[INDEX non_conj_sement,XARG #agr]]].
  3. canonical_synsem := expressed_synsem & canonical_or_unexpressed.
  4. expressed_synsem := synsem.
  5. canonical_or_unexpressed := synsem_min0.
  6. synsem_min0 := synsem_min & [LOCAL mod_local,NONLOC non-local_min].

Which ends with a bunch of basic setup types except for constraining for relation for an adjective to be ‘basically adjectivally’ on the first two lines.  Also on these first two lines, it specifies that its subject and its specifier, if any, must be completed (i.e., empty) and agree with its non-conjunctive argument (which is not to say that it cannot be conjunctive, but that it modifies the conjunction as a whole, if so.)  Whether or not it is expressed will determine if there are any further predicates about its arguments or if its unexpressed argument is identified by an otherwise unreferenced variable in any resulting MRS.

The lexical grounding of this type specification is given below, indicating that it may (or not) have phonology (e.g., pronunciation, such as whether its onset is voiced) and if and how and with what punctuation it may appear, if any.  In general, a semantic argument may be lexical or phrasal and optional but if it appears it corresponds to some semantic index (think variable) in sort of predicate in any resulting MRS.  (The *_min types do not constrain the values of their features any further).

  1. basic_lex_synsem := abstr_lex_synsem & lex_or_nonlex_synsem.
  2. abstr_lex_synsem := canonical_lex_or_phrase_synsem & [LKEYS lexkeys].
  3. canonical_lex_or_phrase_synsem := canonical_synsem & lex_or_phrase.
  4. lex_or_phrase := synsem_min2.
  5. synsem_min2 := synsem_min1 & [LEX luk,MODIFD xmod_min,PHON phon_min,PUNCT punctuation_min].
  6. synsem_min1 := synsem_min0 & [OPT bool,–MIN predsort,–SIND *top*].
  7. adj_synsem_lex_or_phrase := basic_adj_synsem_lex_or_phrase &[LOCAL[CAT.HEAD.MOD <synsem_min &[LOCAL[CAT[HEAD basic_nom_or_ttl & [POSS -],VAL[SUBJ <>,SPR.FIRST synsem &[–MIN quant_or_deg_rel],COMPS <>],MC na],CONJ cnil],–SIND #ind]>,CONT.HOOK.XARG #ind]].

Note that an adjective is not possess-able and that it modifies something nominal (or a title) and that if it has a specifier that it is a quantifier or degree (e.g., ‘very’).  Again, an adjective cannot function as a main clause or be conjunctive (in and of itself).

Finally, if you look far above you will see that the basic semantics of an adjective with an additional semantic argument is ‘intersective’, as in:

  1. isect_synsem := abstr_lex_synsem & [LOCAL[CAT.HEAD.MOD <[LOCAL intersective_mod,NONLOC.REL 0-dlist]>,CONT.HOOK.LTOP #hand],LKEYS.KEYREL.LBL #hand].

Here, the length 0 difference list and the following definitions indicate that intersective semantics do not accept anything but local modification:

  1. intersective_mod := mod_local.
  2. mod_local := *avm*.

AVM stands for ‘attribute value matrix’, which is the structure by which types and their features are defined (with nesting and unification constraints using # to indicate equality).

By now you’re probably getting the idea that there is fairly significant model of the English language, including its lexical and syntactic aspects, but if you look there is a lot about semantics here, too.

Posted on

Is Freebase worth much?

There has been some speculation that Freebase is a vehicle for Metaweb to prosper from its semantic web infrastructure when used for commercial purposes.  As I recall, Metaweb raised over $40 million in Series B around the time they started building Freebase. The investment was led by Goldman Sachs.  Metaweb’s seasoned investors were unlikely to invest so much in a business that cannot project a return on that investment.  Almost certainly, Metaweb has firm plans for realizing over $100 million in revenues.  Most likely, for these investors and the amount of capital, target revenues by 2014, five years after the second round, would be in the vicinity of $1 billion.  Obviously, there is a lot of work to get there from around zero today.

Some of the bubble in raising those funds has burst.  The economy would crimp the valuation and investment if made today.  And the semantic web has yet to produce a winner, so with less enthusiasm, the investment would again be less favorable today.  All this is modulo the business plan.  If the business plan withstands scrutiny and the rate of return from credibly achievable projections justifies investment, they could get the money again, even now.  But no one that I have heard or read over the past few years can explain the business plan adequately – that is, concretely.  I would appreciate any insights or opinions on the topic.  I believe these are smart people, in the company and among its investors, so I am sure it is there.  I just don’t believe in the “we’ll figure out how to make money eventually” business plan in this case.

Some Freebase terms that are worth knowing but are commercially reasonable for any site that provides a free service include:

  1. The terms of service are subject to change (upon posting).
  2. The service may be changed or discontinued at any time and without notice.
  3. Limits concerning access to or use of the services may be established.
  4. Any disputes shall be heard in San Francisco and governed by California law.

Continue reading “Is Freebase worth much?”

Posted on

Google follows Microsoft’s lead towards intelligence

Being a fan of increased intelligence on the web, including Bing’s use of Powerset and True Knowledge, I enjoyed cnet’s report, “Google search gets answer highlights and events.”

Google now shows the following “The Empire State Building rises to 1250 ft (381 m) at the 102nd floor” in response to the classic semantic web test question.

Also, Google leverages more of the content of text or structure of linked data in its Rich Snippet answers:

Rich Snippet shows Google "understands" events

As search engines increase their understanding of concepts and how to extract them from content or linked data and present them as Google does here or above in a sentence, the web will begin to feel a lot smarter. 

As these simple enhancements indicate, the intelligent web is taking off and that feeling of intelligence will come sooner than expected.  Of course, there is a long way to go.   For more on that, I here there is an upcoming issue of AI Magazine that will survey the state of the art in question answering, including coverage of Vulcan’s Project Halo and IBM’s Jeopardy effort, among others.  Also, if you are interested in what bright minds are looking forward to in this regard, see Nova Spivak’s recent blogging and his post on “will the web become conscious?”

Posted on

A Common Upper Ontology for Advanced Placement tests

I have previously written about the lack of a common upper ontology in the semantic web and commercial software markets (e.g., business rules).  For example, the lack of understanding of time limits the intelligence and ease of use of software in business process management (BPM) and complex event processing (CEP).  The lack of understanding of money limits the intelligence and utility of business rules management systems (BRMS) in financial services and the capital markets.   And, more fundamentally, understanding time and money (among other things, such as location, which includes distance) requires a core understanding of amounts.

The core principle here is that software needs to have a common core of understanding that makes sense to most people and across almost every application.  These are the concepts of Pareto’s 80/20 Principle.  A concept like building could easily be out, but concepts like money and time (and whatever it takes to really understand money and time) are in.  Location, including distance, is in.  Luminousity could be out, but probably not if color is in.  Charge and current could be out, but not if electricity or magnetism is in.  The cutoff is less scientific than practical, but what is in has to be deeply consistent and completely rational (i.e., logically rigorous).[2] Continue reading “A Common Upper Ontology for Advanced Placement tests”

Posted on

Over $100m in 12 months backs natural language for the semantic web

Radar Networks is accelerating down the path towards the world’s largest body of knowledge about what people care about using Twine to organize their bookmarks.  Unlike social bookmarking sites, Twine uses natural language processing technology to read and categorize people’s bookmarks in a substantial ontology.  Using this ontology, Twine not only organizes their bookmarks intelligently but also facilitates social networking and collaborative filtering that result in more relevant suggestions of others’ bookmarks than other social bookmarking sites can provide.

Twine should rapidly eclipse social bookmarking sites, like Digg and Redditt.  This is no small feat!

The underlying capabilities of Twine present Radar Networks with many other opportunities, too.  Twine could spider out from bookmarks and become a general competitor to Google, as Powerset hopes to become.  Twine could become the semantic web’s Wikipedia, to which Metaweb’s Freebase aspires. Continue reading “Over $100m in 12 months backs natural language for the semantic web”

Posted on

The $50 Business Rule

Work on acquiring knowledge about science has estimated the cost of encoding knowledge in question answering or problem solving systems at $10,000 per page of relevant textbooks. Regrettably, such estimates are also consistent with the commercial experience of many business rules adopters. The cost of capturing and automating hundreds or thousands of business rules is typically several hundred dollars per rule. The labor costs alone for a implementing several hundred rules too often exceed $100,000.

The fact that most rule adopters face costs exceeding $200 per rule is even more discouraging when this cost does not include the cost of eliciting or harvesting functional requirements or policies but is just the cost of translating such content into the more technical expressions understood by business rules management systems (BRMS) or business rule engines (BRE).

I recommend against adopting any business rule approach that cannot limit the cost of automating elicited or harvested content to less than $100 per rule given a few hundred rules. In fact, Automata provides fixed price services consistent with the following graph using an approach similar to the one I developed at Haley Systems.

Cost per Harvested or Elicited Rule

Continue reading “The $50 Business Rule”