INFORMATION RETRIEVAL USING NEURAL NETWORKS VINEETH REDDY ANUGU - PowerPoint PPT Presentation

INFORMATION RETRIEVAL USING NEURAL NETWORKS VINEETH REDDY ANUGU CMSC 676 INFORMATION RETRIEVAL

INTRODUCTION Ø How can we use neural networks for Information Retrieval? Ø Neural Networks is a field of machine learning Ø How can we train a machine learning model and use that for specific application of information retrieval? Ø This process is highly data demanding

WHAT ARE NEURAL NETWORKS (CNN)? Ø Neural networks learn from raw data Ø Supervised and Unsupervised learning Ø CNNs are set of filters that extract patterns in data Ø Used for image recognition, can be used on textual data Ø CNNs are feed-forward networks

RELATED WORK

THE DEEP RELEVANCE MATCHING MODEL(DRMM) Ø Used for ad-hoc retrieval by relevance matching Ø Convolutional Neural Network is used in this model Ø DRMM is an interaction-based model Ø Interaction-based models looks for interaction between queries and terms Ø Requires Matching Histogram Mapping as input Ø Involves a term gating network

DRMM ARCHITECTURE

WORD2VEC MODEL Ø Used to learn embeddings of given data Ø Embeddings are vector representations of data Ø Useful in recommendation systems, embeddings are drawn from user data Ø Two models Ø Skip-gram Model Ø Continuous bag-of-words Model(CBOW)

COMPARE AND CONTRAST

COMPARING DRMM MODEL Ø Two datasets, Robust04 and ClueWeb-09-Cat-B, are used. Ø Comparing DRMM with well established IR Models

COMPARISON ON ROBUST-04

COMPARISON ON CLUEWEB-09-CAT-B

COMPARING WORD2VEC MODEL Ø Comparing the Word2Vec model with the FastText model Ø FastText obtains n-grams instead of embeddings Ø Useful for rare and out-of-vocabulary word embeddings Ø Uses differ based on the task

THANK YOU