Online Semantic Extraction by Backpropagation Neural Network with Various Syntactic Structure Representations Heidi H. T. Yeung City University of Hong Kong Tat Chee Road, Kowloon Tong, Kowloon, Hong Kong, China Summary of Demonstration The sub-symbolic approach on Natural Language Processing (NLP) is one of the mainstreams in Artificial Intelligence. Indeed, we have plenty of algorithms for variations of NLP such as syntactic structure representation or lexicon classification theoretically. The goal of these researches is obviously for developing a hybrid architecture which can process natural language as what human does. Thus, we propose an online intelligent system to extract the semantics (utterance interpretation) by applying a 3-layer back propagation neural network to classify the encoded syntactic structures into corresponding semantic frame types (e.g. AGENT_ACTION_PATIENT). The results are generated dynamically according to training sets and user inputs in webpage-form. It can diminish the manipulating time while using extra tools and share the statistical results with colleagues in clear and standard forms. Functionalities First, a three layer back-propagation neural network (BPNN) is trained with a set of training sentences and expected semantic frame types as the procedures shown in the flowchart below. The training performance and the classification results of both training and testing sentences are listed systematically and graphically. Select the syntactic structure representation (RAAM, SRAAM, NLC-TPR) Initialize the grammatical settings (loading parsing rules, categories, vocabularies, semantic frame types) Implementation We use three connectionist-based encoders to represent the syntactic structures: Recursive Auto-Association Memories (RAAM) [Pollack 1990], Sequential RAAM [Kwasny 1993] and Tensor product representation with Non-linear Compression (our proposed representational algorithm). There is a main program “NLP” which includes all procedures in the semantic extraction and training processes. The flows of semantic extraction procedures are based on Tsang’s and Wong’s (2002) design. Copyright © 2004, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. Load the training sentences (e.g. The man feeds the cat) YES if Representation is NLC-T P NO This intelligent system is implemented with JAVA which is an object-oriented and multiplatform programming language. During the development stage, the programs can be simply tested in console mode with using IDE provided by SUN. Later, the well-designed Java Service Page (JSP) as the interactive user interface to link those classes as Java Beans with the support of Apache Tomcat server. Set the recurrent neural netw ork Parameters Set sub type of the NLC-TPR Set the back propagation neural netw ork Parameters Start encoding and training the BPNN Display the classifying error, the classifying RMS error during training BPNN and the classification results Figure 1 Flowchart of training the BPNN 1042 INTELLIGENT SYSTEMS DEMONSTRATIONS To test the classifying performance of BPNN, user can input the grammatical sentence with the interactive interface provided. The interface will result the plots of the output layer activation of BPNN and the determined semantic frame type. If the expected type is chosen, the classification error is also calculated as a reference. After the training of BPNN, the latest data will be stored as the references and the 4 types of results can be revised from the system. 1) Encoding Progress – Real time plots of the Rootmean-square (RMS) Error of the recurrent NN during training for RAAM and SRAAM 2) Classifying Progress – Real time plots of the RMS Error of the BPNN during training for different data sets and syntactic structure representations. Figure 3 Output layer activation of BPNN The tree from is previewed with the last Flash movie in Figure 4. 3) Cluster Results – The syntactic structure representations and the BPNN hidden activation of training sentences are clustered in the distance map as Postscript files for different data sets and syntactic structure representations. 4) Classifying Results – The output layer activations of BPNN are tabled against both training and testing sentences for different data sets and syntactic structure representations. Special Features Making use of the dynamic properties in Macromedia Flash Movie, we create 3 SWF documents for graphical display. One of them is for plotting the Root-mean square errors in output layers of recurrent neural network for RAAM and SRAAM and that of BPNN during training shown in Figure 2. Figure 4 Tree display The cluster results of memories and BPNN hidden layer activations are generated and stored in server side. Then, a Java class invokes the Cluster Application and Postscript Generating Program in operating system. References Kwasny, S. C., Kalman, B. L., & Chang, N. 1993. Distributed Patterns as Hierarchical Structures. Proceedings of the World Congress on Neural Networks, Portland, OR, July 1993, v. II, pp. 198-201. Pollack, J. B. 1990. Recursive distributed representations. Artificial Intelligence, 36, pp. 77-105. Tsang, W. M., and Wong, C. K. 2002. Extracting Frame Based Semantics from Recursive Distributed Representation - A Connectionist Approach to NLP. IC-AI' 02. This System is powered by Apache Tomcat 4.1, Macromedia Flash MX Figure 2 Classifying progress display The second one is the activation plots for BPNN output layer activation as below. Copyright © 2004, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. INTELLIGENT SYSTEMS DEMONSTRATIONS 1043