1 / 38

An IR Approach to Multi-document Summarization

An IR Approach to Multi-document Summarization. Joint work with Prasad Pingali, Jagadeesh J, Chandan Kumar and Praveen B. Vasudeva Varma, IIIT Hyderabad. Agenda. Multi-document Summarization (MDS) - Motivation Three Flavors of MDS Query focused Summarization Update Summarization

seddy
Download Presentation

An IR Approach to Multi-document Summarization

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. An IR Approach to Multi-document Summarization • Joint work with Prasad Pingali, Jagadeesh J, Chandan Kumar and Praveen B Vasudeva Varma, IIIT Hyderabad

  2. Agenda • Multi-document Summarization (MDS) - Motivation • Three Flavors of MDS • Query focused Summarization • Update Summarization • Personalized summaries • Abstract Summarization – A demo

  3. Motivation

  4. Summaries Can Help !

  5. DUC Task Description • Task is to create, from the document set, a response which answers the information need expressed • Input: Information need, Cluster of relevant documents (assumed to contain answer) • Output: Answer of required length

  6. Extract vs. Abstract Summarization • We conducted a study (2005) • Generated best possible extracts • Calculated the scores for these extracts • Evaluation with respect to the reference summaries

  7. Sentence Extraction Approaches • MEAD: sentence-level and inter-sentence features, cluster centroids, position, TF*IDF • CLASSY: Learned HMM model to identify summary and non-summary sentence, QR algorithm, linguistic component • Mani & Erkon: Graph-connectivity model • Lin & Hovy: sentence position, term frequency, topic signature and term clustering • LexPageRank, Hardy, Harabagiu and Lacatusu etc….. • Supervised Approaches: sentence classifiers trained using human-generated summaries as training examples for feature extraction and parameter estimation

  8. Our approach: Big Picture

  9. Our approach • Documents should be ranked in order of probability of relevance to the request or information need, as calculated from whatever evidence is available to the system • Central idea: • Query Dependent Ranking (DUC 2005) • Language Models (HAL, RBLM) • Query Independent Ranking (DUC 2006) • Sentence Prior

  10. Relevance Based Language Models (RBLM) • An IR approach • Query and document are samples of unknown relevance model R • Overcomes the problem of sparseness of the document language model • Conditional Sampling computes the conditional probabilities, decomposed from joint probabilities

  11. Hyperspace Analogue to Language (HAL): Co-occurrence strengths • A concept can be understood from its context: Constructs dependencies of a term w on other terms based on their occurrence in its context in the corpus • Input: corpus of documents with vocabulary of size |T | • Output: a |T | x |T | matrix, rows represent the vectors corresponding to different terms in vocabulary • Process: Sliding window of K words • Words in window are proportional with strengths inversely proportional to their distance • Vectors of semantically related words are similar • Frequency counts are converted into probabilities ─ P(w1/w2)

  12. Putting All Together HAL Feature:

  13. Modified HAL - Addressing Phrases • Phrases are identified using Chunker • Co-occurrence strength of word on chunks • Word/Phrase is weighted proportional to their TFIDF value

  14. Recap

  15. Query independent Features: Sentence Prior • Captures importance of sentence explicitly using pseudo relevant documents (Web, Wikipedia, DUC Document Sets) • Based on Domain knowledge, Background Information, Centrality • Log Linear Relevance • Information Measure in a sentence • Entropy is a measure of information contained in a message

  16. Recap

  17. Summary Generation

  18. Evaluation: DUC 2005

  19. Analysis: Individual Feature Performance

  20. Parameter Selection Based on DUC 2005 data

  21. DUC 2006: Official Results Total 38 systems participated Significant difference between first two systems 5th Rank on linguistic quality

  22. An Application: Ask Budha

  23. Progressive Summarization • Emerging area of research in summarization • Summarization with a sense of prior knowledge • Introduced as “Update Summarization” at DUC 2007, TAC 2008, TAC 2009 • Generate a short summary of a set of newswire articles, under the assumption that the user has already read a given set of earlier articles. • To keep track of news stories, reviews of products

  24. Key challenge • To detect information that is not only relevant but also new given the prior knowledge of reader • Relevant and new Vs • Non-Relevant and new Vs • Relevant and redundant

  25. Novelty Detection • Identifying sentences containing new information (Novelty Detection) from cluster of documents is the key of progressive summarization • Shares similarity with Novelty track at TREC from 2002 – 2004 • Task 1: Extract relevant sentences from a set of documents for a topic • Task 2: Eliminate redundant sentences from relevant sentences • Progressive summarization differs, as in producing summary from novel sentences (requires scoring and ranking)

  26. Three level approach to Novelty Detection • Sentence Scoring • Developing new features that capture novelty along with relevance of a sentence • Ranking • Sentences are re ranked based on the amount of novelty it contains • Summary Generation • Summary of recent cluster will have minimum overlap with previous summaries • A selected pool of sentences that contain novel facts. All remaining sentences are filtered out

  27. Consider a stream of articles published on a topic over time period T All articles published from time 0 to t are considered to be read previously (prior Knowledge) Articles published from t to T are new that contains new information. Let td represent the chronological time stamp of document d. Novelty Detection Features: NF (Novelty Factor) and New Words

  28. Sentence Scoring Features

  29. Ranking Features • Ranked set is Re-ordered based on redundancy score of each sentence • Re-Ranked set used for summary generation • Cosine Similarity (CoSim) • Each sentence in new cluster compared against all sentences in previous clusters • Average Cosine similarity value is considered as the redundancy score Rank = relweight*rank - redweight*redundancy_score relweight = 0.9, redweight = 1-relweight

  30. Evaluation and Results • TAC 2008 Update Summarization data for training: 48 topics • Each topic divided into A, B with 10 documents • Summary for cluster A is normal summary and cluster B is update summary • TAC 2009 update Summarization for testing: 44 topics • Baseline summarizer generates summary by picking first 100 words of last document • Run1 – DFS + SL1 • Run2 – PHAL + KL

  31. Summary Generation

  32. Personalized Summarization - Motivation • When different humans summarize the same text • 71% overlap [Marcu-1997] • 25% overlap [Rath-1961] • 46% overlap [Salton-1997] • They include different content from each other, reflecting their personal interest and background knowledge • Each User has different perspective on the same text • Need to incorporate user in the automatic summarization process • Summarization is not only a function of the input text but also of its reader.

  33. Term interest differ for each person • Incorporating User Model P(w/Mu) to smooth the original Document distribution

  34. Experiments and Evaluation • Users: Research scholars from different fields of computer science • Web-based profile creation: Personal information available on web- a conference page, a project page, an online paper, or even in a Weblog. • Put the person's full name to a search engine ("Vasudeva Varma") and retrieve top 'n' documents to build user profile • Estimate Model P(w/Mu) to incorporate user in sentence extraction process • 5 Users, 25-Doc Clusters • Two versions of summary generated for each user • Generic Summary • Personalize summary • Each User was asked to asked to give his relevance score to the summary on a 5-point scale..

  35. Evaluation Average Scores for different Uses Scores for different topics for a user

  36. Abstract Summarization – A Demo • http://compare.setooz.com • Alias http://comparison.to

  37. Thank you Questions?

More Related