1/13
Introduction
NLP and Transformers
Networked Mathematics
Networked Mathematics™
NLP tools for Better Science
Valeria de Paiva
Topos Institute
April 2022
Valeria de Paiva Topos
2/13
Introduction
NLP and Transformers
Networked Mathematics
Thanks, Fabio and Davide!
Valeria de Paiva Topos
3/13
Introduction
NLP and Transformers
Networked Mathematics
NLP research?
Valeria de Paiva Topos
4/13
Introduction
NLP and Transformers
Networked Mathematics
Semantics & Language
Samsung SRA (2019): Dialogue and Knowledge
Representation Lab,
project: systems to make Bixby (voice personal assistant)
communicate well with home appliances at SmartHome
Nuance Comms (2012-2018): AI to make sounds into
knowledge, health systems, automotive, law, CRM, banks,
insurance, etc
projects: personal assistant for Living Room (TV 2nd screen),
PA for automotive companies
Rearden Commerce (2011-2012): a white-labelling shop for
travel and expenses/procurement systems. air travel tickets,
hotels, shows & sports, restaurants, parking, etc
project: a Groupon-like app, ontologies, discover what hotel
reviewers really valued
Cuil (2008-2010) search analysis, PARC Forum: Adventures in
Searchland
Valeria de Paiva Topos
5/13
Introduction
NLP and Transformers
Networked Mathematics
How to think about language understanding?
Valeria de Paiva Topos
6/13
Introduction
NLP and Transformers
Networked Mathematics
How to think about language understanding?
Valeria de Paiva Topos
7/13
Introduction
NLP and Transformers
Networked Mathematics
Conversational UI applications
AI Summit 2018, Luxembourg
Valeria de Paiva Topos
8/13
Introduction
NLP and Transformers
Networked Mathematics
Conversational Reasoning Engines
AI Summit 2018, Luxembourg
Valeria de Paiva Topos
9/13
Introduction
NLP and Transformers
Networked Mathematics
Can we read Math texts?
Work of almost nine years at PARC was out of reach when I
left in 2008
Pleased to report that almost all of it is now available
open-source, redone from scratch, using new techniques.
work of Katerina Kalouli, PhD thesis 2021
Hy-NLI: a hybrid NLI engine, computes inference in a hybrid
way by employing the symbolic system GKR4NLI and the deep
learning model BERT;
XplaiNLI: explainability of the Hy-NLI system and sketches of
explanations for the decisions made by each component of
Hy-NLI;
GKR4NLI a symbolic NLI engine that computes the inference
relation between a pair of sentences
GKR parser transforms a given sentence into a layered
semantic graph, its symbolic representation
Demos for all components!
Valeria de Paiva Topos
10/13
Introduction
NLP and Transformers
Networked Mathematics
Knowledge Graph for Maths?
Doing semantics is expensive, we can get much information
with less than full semantics
But we need the specific vocabulary. How do we get it?
Term extractors: Open Tapioca, TextRank, DyGIE++,
Parmenides (our NIST collaborators) the ones we tried
No one is specific for mathematics, but DyGIE++ is about
scientific text
Bottleneck is always evaluation: difficult and not super fun
Human annotation in this domain is expensive and difficult to
obtain
Can we use ‘silver datasets’ instead?
Valeria de Paiva Topos
11/13
Introduction
NLP and Transformers
Networked Mathematics
Knowledge Graph for Maths
Author keywords are ‘guaranteed’ sound keywords, but won’t
be all the story
Need to add ‘trivial’ keywords for experts – these are
extractive, e.g. ‘category’ in a journal about Category Theory
there are many expressions that are keyword-like, e.g. ”Future
work”, that do not correspond to concepts
So ‘linguistic concepts’ are broader than mathematical
concepts
Can training on guaranteed keywords be enough to obtain
concepts and only concepts?
Valeria de Paiva Topos
12/13
Introduction
NLP and Transformers
Networked Mathematics
Conversational UI applications
AI Summit 2018, Luxembourg
eventually Hy-NLI!
Valeria de Paiva Topos
13/13
Introduction
NLP and Transformers
Networked Mathematics
Conclusions
Knowledge representation for Mathematics is fun, interesting,
useful. and underdeveloped!
Payoff is incredibly high, see our blog post ‘Introducing the
MathFoldr Project’ https://topos.site/blog/2021/07/
introducing-the-mathfoldr-project/
NLP tools are becoming better every week, need to make use
of this productivity
Valeria de Paiva Topos

Networked Mathematics: NLP tools for Better Science

  • 1.
    1/13 Introduction NLP and Transformers NetworkedMathematics Networked Mathematics™ NLP tools for Better Science Valeria de Paiva Topos Institute April 2022 Valeria de Paiva Topos
  • 2.
    2/13 Introduction NLP and Transformers NetworkedMathematics Thanks, Fabio and Davide! Valeria de Paiva Topos
  • 3.
    3/13 Introduction NLP and Transformers NetworkedMathematics NLP research? Valeria de Paiva Topos
  • 4.
    4/13 Introduction NLP and Transformers NetworkedMathematics Semantics & Language Samsung SRA (2019): Dialogue and Knowledge Representation Lab, project: systems to make Bixby (voice personal assistant) communicate well with home appliances at SmartHome Nuance Comms (2012-2018): AI to make sounds into knowledge, health systems, automotive, law, CRM, banks, insurance, etc projects: personal assistant for Living Room (TV 2nd screen), PA for automotive companies Rearden Commerce (2011-2012): a white-labelling shop for travel and expenses/procurement systems. air travel tickets, hotels, shows & sports, restaurants, parking, etc project: a Groupon-like app, ontologies, discover what hotel reviewers really valued Cuil (2008-2010) search analysis, PARC Forum: Adventures in Searchland Valeria de Paiva Topos
  • 5.
    5/13 Introduction NLP and Transformers NetworkedMathematics How to think about language understanding? Valeria de Paiva Topos
  • 6.
    6/13 Introduction NLP and Transformers NetworkedMathematics How to think about language understanding? Valeria de Paiva Topos
  • 7.
    7/13 Introduction NLP and Transformers NetworkedMathematics Conversational UI applications AI Summit 2018, Luxembourg Valeria de Paiva Topos
  • 8.
    8/13 Introduction NLP and Transformers NetworkedMathematics Conversational Reasoning Engines AI Summit 2018, Luxembourg Valeria de Paiva Topos
  • 9.
    9/13 Introduction NLP and Transformers NetworkedMathematics Can we read Math texts? Work of almost nine years at PARC was out of reach when I left in 2008 Pleased to report that almost all of it is now available open-source, redone from scratch, using new techniques. work of Katerina Kalouli, PhD thesis 2021 Hy-NLI: a hybrid NLI engine, computes inference in a hybrid way by employing the symbolic system GKR4NLI and the deep learning model BERT; XplaiNLI: explainability of the Hy-NLI system and sketches of explanations for the decisions made by each component of Hy-NLI; GKR4NLI a symbolic NLI engine that computes the inference relation between a pair of sentences GKR parser transforms a given sentence into a layered semantic graph, its symbolic representation Demos for all components! Valeria de Paiva Topos
  • 10.
    10/13 Introduction NLP and Transformers NetworkedMathematics Knowledge Graph for Maths? Doing semantics is expensive, we can get much information with less than full semantics But we need the specific vocabulary. How do we get it? Term extractors: Open Tapioca, TextRank, DyGIE++, Parmenides (our NIST collaborators) the ones we tried No one is specific for mathematics, but DyGIE++ is about scientific text Bottleneck is always evaluation: difficult and not super fun Human annotation in this domain is expensive and difficult to obtain Can we use ‘silver datasets’ instead? Valeria de Paiva Topos
  • 11.
    11/13 Introduction NLP and Transformers NetworkedMathematics Knowledge Graph for Maths Author keywords are ‘guaranteed’ sound keywords, but won’t be all the story Need to add ‘trivial’ keywords for experts – these are extractive, e.g. ‘category’ in a journal about Category Theory there are many expressions that are keyword-like, e.g. ”Future work”, that do not correspond to concepts So ‘linguistic concepts’ are broader than mathematical concepts Can training on guaranteed keywords be enough to obtain concepts and only concepts? Valeria de Paiva Topos
  • 12.
    12/13 Introduction NLP and Transformers NetworkedMathematics Conversational UI applications AI Summit 2018, Luxembourg eventually Hy-NLI! Valeria de Paiva Topos
  • 13.
    13/13 Introduction NLP and Transformers NetworkedMathematics Conclusions Knowledge representation for Mathematics is fun, interesting, useful. and underdeveloped! Payoff is incredibly high, see our blog post ‘Introducing the MathFoldr Project’ https://topos.site/blog/2021/07/ introducing-the-mathfoldr-project/ NLP tools are becoming better every week, need to make use of this productivity Valeria de Paiva Topos