Contents Introduction Outline of our approach Background o Tf-Idf ranker o Semantic similarity between sentences Details of our approach Results Conclusion References
Introduction Traditional information retrieval systems are particularly susceptible to all the problems posed by the richness of natural language. In particular multitude of ways in which the same concepts can be described. Overall context of the user input and the document is ignored. Traditional TF IDF Ranker ignores the relatedness of concepts. Searches for the exact word match. Introduction of semantic analyzer will improve the performance.
Introduction (cont..) Aim of the project is to use traditional TF IDF ranker along with semantic analyzer to retrieve the documents. And to compare the performance of the new system with the traditional tf idf ranker.
Introduction (cont..) This project uses, o Text Retrieval Conference (TREC) data set named Confusion track for validation . o Wordnet lexical database o.NET framework (wordnet.net)
Input Query Documents Primary filter TF IDF Ranker Pre-processor Doc ID, Weight pairs Traditional TF IDF Ranker Document s Final Docs Outline of our approach
Input Query Document s Primary filter TF IDF Ranker Pre-processor Doc ID, Weight pairs TF IDF Ranker with introduction of Semantic knowledge Document s Final Docs Semantic similarity Outline of our approach (cont..)
Input Query Documents TF-IDF Ranker II Wordnet semantic Analyzer Pre-processor Doc ID, Semantic score DocID, Keywords Final Docs Corpus Word,DF pairs Find the Keywords from each doc Use Tf and Df (use Corpus) Outline of our approach (cont..) Docs got from traditional tf idf approach
Pre-processor Tokenize Remove stop words Outline of our approach (cont..)
Background Tf-Idf ranker : Tf-idf ranker is used as a weighting factor in information retrieval and text mining. Terms that appear often in a document should get high weights. The more often a document contains a term, the more likely that the document is about the term. It is captures using Term frequency ( TF ). Terms that appear in many documents should get a low weight, which is captured using Inverse Document Frequency ( IDF ). The weight of a term in a document is calculated using below formula , W i,j =TF i,j * log (N/DF i )
Semantic similarity between sentences : Semantic similarity between sentences is calculated using semantic information and the word order information. This project has used an implementation which calculates the semantic relatedness between two set of strings. The implementation uses Wordnet lexical database, to calculate the semantic relatedness. The score lies between 0 and 1. 0 representing least similarity score. 1being highest.
Wordnet : Wordnet is the product of a research project at Princeton University . Information in Wordnet is organized around logical groupings called synsets. Each synset consists of a list of synonymous word forms and semantic pointers that describe relationships between the current synset and other synsets. In Wordnet, each part of speech words (nouns/verbs...) are organized into taxonomies where each node is a set of synonyms (synset) represented in one sense.
Wordnet (cont..) If a word has more than one sense, it will appear in multiple synsets at various locations in the taxonomy. Wordnet defines relations between synsets and relations between word senses. A relation between synsets is a semantic relation, and a relation between word senses is a lexical relation.
Wordnet (cont..) For example, The shortest path between male and female in Fig. 1 is male-person- female, the minimum path length is 2. The minimum path length between female and teacher is 5.
Details of our approach Traditional TF-IDF Ranker Step1:Preprocess input query o Tokenization o Remove stop words Step2: Apply Tf-Idf ranker TF-Idf ranker would identify number of times each word appears in each of the documents as shown below. Where TFij is the term frequency of word wi in document Dj. DF i indicates document frequency of word Wi in document collection D1D2D3, DNDN DF W1TF11TF12 TF1NDF1 W2TF21TF22 TF2NDF2 W3TF31TF32 TF3NDF3 : : WnTFn1TFn2 TFnNDFn
Details of our approach (cont..) Calculating the weight: The weight of each word is calculated using below formula. W i,j =TF i,j * log (N/DF i ) D1D2D3, DNDN DF W1W 11 W 12 W 1N DF1 W2W 21 W 22 W 2N DF2 W3W 31 W 32 W 3N DF3 : : WnW n1 W n2 W nN DFn Weight sumS1S2 SNSN
Details of our approach (cont) Step3 : Retrieve the documents Sort all the documents according to the weights. Pick top Q documents for further processing. Q is chosen such as the weight of each document crosses a particular threshold . Improved TF-IDF Ranker Step1: We choose top S from the step3 of previous method. Here we use another threshold to get the set of docs for further processing. Step2: Extract the keywords (Words which have high TF and low DF) from each document. DocDFWeight W1TF1DF1We1 W2TF2DF2We2 W3TF3DF3We3 : : WnTFnDFnWen
Details of our approach (cont) Corpus containing IDF (logN/DF) of each word from docs
Details of our approach (cont..) Step 3: For each document, calculate the semantic similarity score between its keyword set and the input query. Step 4: Sort the docs w.r.t to score. Eliminate the docs with score less than a specified threshold Step 5: Display the docs.
Confusion Track result set Results
Results: Old system vs New system Results (cont..)
Calculating precision & recall for 10 queries Results (cont..)
Precision& Recall bar chat: Old system vs New system Results (cont..)
Screenshots Traditional IF IDF Ranker
Screenshots (cont..) Improved IF IDF Ranker(with semantic knowledge)
Conclusion This project has improvised traditional TF-IDF ranker by introducing Semantic analyzer. Successfully showed that, using semantic analyzer has good precision and recall values. Next, it used a dataset from Text Retrieval Conference Data (TREC) to validate the project. One limitation of Tf-Idf Ranker is, terms that occur in query input text but that cannot be found in documents gets zero scores.
References  R. Rada, H. Mili, E. Bichnell, and M. Blettner, “Development and Application of a Metric on Semantic Nets,” IEEE Trans. System, Man, and Cybernetics, vol. 9, no. 1, pp ,  Li, Yuhua,et.al, “Sentence Similarity Based on Semantic Nets and Corpus Statistics,” IEEE Trans on knowledge and data engineering, vol 18, no.8,2006.  Dao, Thanh, Troy Simpson, “Measuring similarity between the sentences”.Web.  R. Richardson, A. F. Smeaton and J. Murphy, “Using WordNet as a Knowledge Base for Measuring Semantic Similarity between Words,” School of Computer Applications, Dublin City University.Web.  TfIdf Ranker, ‘http://vetsky.narod2.ru/catalog/tfidf_ranker/’.web.  Confusion track, TREC dataset ‘http://trec.nist.gov/data/t5_confusion.html’.Web.