Presentation on theme: "Connectionist Sentence Comprehension and Production System A model by Dr. Douglas Rohde, M.I.T by Dave Cooke Nov. 6, 2004."— Presentation transcript:
Connectionist Sentence Comprehension and Production System A model by Dr. Douglas Rohde, M.I.T by Dave Cooke Nov. 6, 2004
Overview Introduction –A brief overview of Artificial Neural Networks –The basic architecture Introduce Douglas Rohde's CSCP model –Overview –Penglish Language –Architecture –Semantic System –Comprehension, Prediction, and Production System –Training –Testing –Conclusions Bibliography
A Brief Overview oBasic definition of an Artificial Neural Network oA network of interconnected “neurons” inspired by the biological nervous system. oThe function of an Artificial Neural Network is to produce an output pattern from a given input. oFirst described by Warren McCulloch and Walter Pitts in 1943 in their seminal paper “A logical calculus of ideas imminent in nervous activity”.
Artificial neurons are modeled after biological neurons The architecture of an Artificial Neuron
Architecture -- Structure oNetwork Structure oMany types of neural network structures oEx Feedforward, Recurrent oFeedforward oCan be single layered or multi-layered oInputs are propagated forward to the output layer
Architecture -- Recurrent NN oRecurrent Neural Networks o Operate on an input space and an internal state space – they have memory. o Primary types of Recurrent neural networks osimple recurrent ofully recurrent oBelow is an example of a simple recurrent network (SRN)
Architecture -- Learning oLearning used in NN's oLearning = change in connection weights oSupervised networks: network is told about correct answer oex. back propagation, back propagation through time, reinforcement learning oUnsupervised networks: network has to find correct input. ocompetitive learning, self-organizing or Kohonen maps
Architecture -- Learning (BPTT) oBackpropagation Through Time (BPTT) is used in the CSCP Model and SRNs oIn BPTT the network runs ALL of its forward passes then performs ALL of the backward passes. oEquivalent to unrolling the network backwards through time
The CSCP Model oConnectionist Sentence Comprehension and Production model oPrimary Goal: learn to comprehend and produce sentences developed in the Penglish( Pseudo English) language. oSecondary Goal: to construct a model that will acount for a wide range of human sentence processing behaviours.
Basic Architecture oA Simple Recurrent NN is used oPenglish (Pseudo English) was used to train and test the model. oConsists of 2 separate parts contected by a “message layer” oSemantic System (Encoding/Decoding System) oCPP system oBackpropagation Through Time (BPTT) is the learning algorithm. omethod for learning temporal tasks
Penglish oGoal: to produce only sentences that are reasonably valid in english oBuilt around the framework of a stochastic context-free grammar. oGiven a SCFG it is easy to generate sentences, parse sentences, and perform optimal prediction oSubset of english some grammatical structures used are o56 verb stems o45 noun stems oadjectives, determiners, adverbs, subordinate clauses oseveral types of logical ambiguity.
Penglish oPenglish sentences do not always sound entirely natural even though constraints to avoid semantic violations were implemented oExample sentences are: o(1) We had played a trumpet for you o(2) A answer involves a nice school. o(3) The new teacher gave me a new book of baseball. o(4) Houses have had something the mother has forgotten
The CSCP Model Semantic System CPP System Start stores all propositions seen for current sentence
Semantic System Propositions loaded sequentially Propositions stored in Memory
Training (SS) oBackpropagation oTrained separate and prior to the rest of the model. oThe decoder: uses standard single-step backpropagation oThe encoder is trained using BPTT. oMajority of the running time is in the decoding stage.
CPP System Error measure Phonologically encoded word. The CPP System
CPP System (cont.) Starts here by trying to predict next word in sentence. Goal to produce next word in sentence and pass it to Word Input Layer
4. BPTT The CPP System - Training 1. BPTT starts here. 2. Backpropagated to here. 3. Previously recorded output errors are injected here
Training o16 Penglish training sets oSet = 250,000 sentences, total = 4 million sentences o50 000 weight updates per set = 1 epoch oTotal of 16 epochs. oThe learning rate start at.2 for the first epoch and then was gradually reduced over the course of learning. oAfter the Semantic System the CPP system was similarily trained oTraining began with limited complexity sentences and complexity increased gradually. oTraining a single network took about 2 days on a 500Mhz alpha. Total training time took about two months. oOverall 3 networks were trained
Testing o50,000 sentences o33.8% of testing sentences also appeared in one of the training sets. oNearly all of the sentences had 1 or 2 propositions. o3 forms of measurement are used in measuring comprehension. omultiple choice measure oReading time measure oGrammaticality rating measure
Testing (Multiple Choice) oExample: “When the owner let go, the dog ran after the mailman.” oExpressed as [ran after, theme, ?] oPossible answers oMailman (correct answer) oowner, dog, girls, cats. (distractors) oError measure is oWhen applying four distractors, the chance performance is 20% correct.
Testing (Reading Time) oAlso known as Simulated Reading Time oIt’s a weighted average of 4 components. o1 and 2 “Measure the degree to which the current word was expected” o3 rd “The change in the message that occurred when the current word was read” o4 th “The average level of activation in the message layer” oThe four components are multiplied by scaling factors to achieve average values of close to 1.0 for each of them and a weighted average is then taken. oRanges from.4 for easy words to 2.5 or more for very hard words.
Testing (Grammaticality) oThe Grammaticality Method o(1) prediction accuracy (PE) oIndicator of syntactic complexity oInvolves the point in the sentence at which the worst two consecutive predictions occur. o(2) comprehension performance (CE) oAverage strict-criterion comprehension error rate on the sentence. oIntented to reflect the degree to which the sentence makes sense. oSimulated ungrammaticality rating (SUR) oSUR = (PE – 8) X (CE + 0.5) ocombines the two components into a single measure of ungrammaticality
Conclusions oGeneral Comprehension Results ofinal networks are able to provide complete, accurate answer oGiven NO choices 77% oGiven 5 choices 92% oSentential Complement Ambiguity oStrict criterion error rate 13.5% oMultiple choice 2% oSubordinate Clause Ambiguity oEx. Although the teacher saw a book was taken in the school. oIntransitive, weak bad, weak good condition, strong bad, and strong good all were under 20% error rate on multiple choice questions.
Bibliography 1.Artificial Intelligence 4 th ed, Luger G.F., Addison Wesley, 2002 2.Artificial Intelligence 2 nd ed, Russel & Norvig, Prentice Hall, 2003 3.Neural Networks 2 nd ed, Picton P., Palgrave, 2000 4.A connectionist model of sentence comprehension and production, Rohde D., MIT, March 2 2002 5.Finding Structure in Time, Elman J.L, UC San Diego, Cognitive Science, 14, 179-211, 1990 6.Fundamentals of Neural Networks, Fausett L, Pearson, 1994