Yingjin Xu School of Philosophy, Fudan University

From Turing to Confucius: philosophical inspirations

underlying different approaches in Natural Language Processing

Yingjin XuSchool of Philosophy, Fudan University

Philosophy and AI

• AI is special among other disciplines in the sense that it does not have an unified theoretical foundation as classical physics does.

• AI systems are often technical realization of some philosophical ideas without some scientific theory as a media between the two.

But

• The Philosophical is even more chaotic than the AI circle.

• Disagreements among philosophers on nearly every level are quite normal. Much worse, no experimental methods can be used to set some criteria.

But how does philosophy evolve?

• 1. By doing arguments.• (common sense plus logic, sometime probability

theory)Intuition plays a role here, comparable to the wanted “friendly-to-user” feature of AI. AI system also needs to send out outputs which are intuitively right to humans. 2. Older theories never die, but they would fade away.

For instance

• Behaviorism both as psychological and philosophical movement has faded away, but may be still alive in a sense.

• Skinner• Watson• Ryle• But why did they fade away?• Cognitive turn and Chomsky

But behaviorism has much impact on NLP

• 1. Turing test idea.• Joseph Weizenbaum:• Eliza chatterbox • Rogerian psychotherapy• The recipe here is: no need to model the inner

language processing structure of the doctor. Only behaviors do matter.

• Is the AERA chatterbox still on this line?

External semantics

As Wittgenstein described in Philosophical Investigation:Meanings of words come from objects stand for them in the reality.

This looks very natural, and it is the core idea of early Wittgenstein.

But

• It has faded away.• It is not refuted in a direct way, but it does not

seem to be promising from the perspective of later Wittgenstein.

• The theory for replacing the former:• Language game theory

However, the external semantic model still has its impacts in AI or NLP

• Instance:• Terry Winograd: SHRDLU• Block world• Philosophical worry:• Not everything represented by words can be

defined as nodes mapped onto external objects.

• Can gestures be defined in this way?

Leibniz: ideal language

• Such a language is composed by:• Linguistic entities clearly representing ideas• And rules which are mechanically computable.

• Ideas are innately fixed, needless to represent external objects.

• Understanding: a mapping job from the natural language to this ideal language.

Instance

• Margaret Masterman: Interlingua-based Machine Translation

• Source language: gusta• Inter language: [CAUSE (X, [BE (Y, PLEASED])])]• Target language: like• But the IL should be very powerful for

representing every possible natural language.

Transfer-based Machine Translation

• SL: Maria me gusta• Syntactic analysis:• gusta[me (Maria)]• gusta[SUBJ(ARG2:NP), OBJ1(ARG1: CASE1)]• gusta[SUBJ(ARG2:NP), OBJ1(ARG1:

CASE1)]→like [SUBJ(ARG1:NP), OBJ1(ARG2:NP)]

Empiricism Vs. Rationalism

• Statistics Vs. Rule-based approach

A Bayesian approach

• An idealized translation is nothing but any mapping job which can make the value of the following formula approach to 1:

• P(target expression/source expression)• In order to do the mapping, the system should try

every candidate target expression and hence compute the result of the foregoing formula, and finally select the most qualified guy out of the candidate pool.

• IBM: Candide system

Kant: Hybrid system

• The basic ideas of the Kantian cognitive architecture:

• Mixture of the Bottom-up and Top-down approach

Instance in NLP

• Sergei Nirenberg and Robert Frederking:• Multi-Engine Machine Translation• Three engines ( one based on statistics, one

based on rules, and one based on instances)• Each one will work on the same source text,

sending out three candidate target texts, them a higher order evaluator will select out the best one.

The philosophical worry

• As the criticism of Kant made by Hegel has shown, the integration would be a big problem for any hybrid approach.

Confucius!

• Confucius would not like rule-based approach.• He never defines concepts or gives axioms.• For instance, he never defines “benevolence”. • Nor would he like statistics or large-scale date-

mining.• Instead, he is a story teller, and stories are not

rules on the one hand, not sufficient for forming big sample space on the other.

长尾真（ Makoto Nagao ）’ s work

• Dual lingua mapping Samples given as below:• I feel dizzy. ↔ 我感到晕。• I feel hungry. ↔ 我感到饿。• He feels hungry. ↔ 他感到饿。• He feels comfortable. ↔ 他感到舒服。• The system would figure out the mapping

relationships on the word level, and once this is done, the grammatical structure will be grasped.

But there is a further problem:

• Embodiment!!!!

• Embodiment is not a discrete module besides the NLP module.

• Instead, full-fledged capacity of NLP has assumed embodiment.

END

• Thank you for your patience!!

• [email protected]• http://jpkc.fudan.edu.cn/s/293/t/608/main.htm