Background Other summarizers Techniques Future improvements Applications Evaluation

Transcription

1 DSV-SU-KTH / Background Other summarizers Techniques Future improvements Applications Evaluation 1 2 is the method where a computer summarizes a text. An extract from a longer original text. A text is given to the computer and it returns a non-redundant shorter text The technique has it s roots in the 60 s. With the Internet and the WWW it has been an awakening interest in summarization techniques. Extraction is much easier than abstraction Abstraction needs understanding and rewriting Most automatic summarization tools make extracts not abstracts Uses original sentences or part of sentences to create abstract projects_full.htm Microsoft Word 97, 98 and Word 2000 have a summarizer for documents. Intelligent Miner for Text -Summarization tool IBM Inxight (XEROX) Corporum Summarizer- Cognit AS (Norway) Pertinence (France) Copernic Summarizer Automated Text Summarization (SUMMARIST) Columbia Newsblaster newsblaster OracleContext Safari web reader for Macintosh under Safari->Services 5 6 1

2 Search engine - extracts in hit lists Business Intelligence- survey news flow Translation - make the text shorter before translating the text Summarize news for SMS, WAP, 3G-format News paper setting and printing Speech synthesis - summarize text before synthesize. Text-To-Speech 7 8 Single-document summarization One document summarized Multi-document summarization Several documents summarized to one Multi-lingual summarization Several documents in several languages to one doc. Multimodal summarization User adapted summary-slanted summary Summarize around some key-concepts Find what the text is about Then decide what so say Then decide how to say it 9 10 SweSum is the first summarizer for Swedish news text Started in 1999 Is available on the Internet Remove all sentences except the first one and the title. Done!

3 Text summarization (extraction) uses statistic, linguistic and heuristic methods A text is divided into sentences Sentence positions (News/Reports) Title words Bold text, Numerical values, Citations Named Entities (Frequency based) 13 Keyword extraction and frequency Language specific Domain specific Also called "open class word nouns, adverbs, adjectives Use morphological information-lemma- ( stemming ) skatt, skatterna, skatten, skattens..=> skatt (tax) öka, ökning, ökningens.. => öka (increase) Key words in news domain 14 User adaptation Utilize user keywords - Obtain slanted summaries Combination function of all rankings with different weights gives the rank of each sentence. Generate all high ranking sentences Voilá the summary! A Client/Server web-based application Web Server HTTP Web Client Summarizer Pass I Tokenizing Scoring Keyword extraction Lexicon 4 Original Text 5 Pass II Sentence Ranking Apache HTTP Server Pass III Summary Extraction HTTP Client (Win Explorer/Netscape/Mac) Summarized Text SweSum Architecture Figure by Nima Mazdak Summarizes Swedish news paper text in HTML/ text format on the WWW. Uses a Swedish key word lexicon that contains words and their possible inflections. During the text summarization are 5-10 key words produced which describes or categorizes the text Key words - A miniature summary words words Inflected version Lemma statsminister statsminister (premier minister) statsministern statsminister statsministerns statsminister statsministrarna statsminister statsministrarnas statsminister..... regeringen regeringen (government) regeringens regeringen regeringarna regeringen regeringarnas regeringen

4 30 percent summarization rate 70 percent compression rate (remove 70 percent of text) Gives 80 percent intact information ~70 percent is redundant information SweSum is available for 10 languages SweSum is available to summarize news texts in Swedish, Danish, Norwegian, English, Spanish, French, Greek, Italian, German and in Farsi (Persian). SweSum has around unique visitors per month A large amount of them are from Spanish, American, Southamerican, Italian and French universities Original text Finnish college gunman kills stm SweSum demo

5 Summarized text The yellow sections are the summarized sections by Swesum, the rest is the original text Pronoun and other anaphor references Kalle sprang. Han sprang fort. Pronoun resolution Clauses can be too long or too short Clause reductions- and clause combination rules Aggregation Analysera mera! Regi: Harold Ramis Medv: Robert De Niro, Billy Crystal, Lisa Kudrow Längd: 1 tim, 45 min Ett av många skäl att glädjas åt Analysera mera är att Robert De Niro här verkligen utövar skådespelarkonst igen. Han accelererar emotionellt från 0 till 100 på ingen tid alls, för att sedan kattmjukt bromsa in och parkera, lugnt och behärskat. Och han är tämligen oemotståndlig. Här har han åstadkommit ännu en intelligent komedi för alla oss vänner av intelligens och komedi, gärna i kombination. SvD Analysera mera! Regi: Harold Ramis Medv: Robert De Niro, Billy Crystal, Lisa Kudrow Längd: 1 tim, 45 min Ett av många skäl att glädjas åt Analysera mera är att Robert De Niro här verkligen utövar skådespelarkonst igen. Robert accelererar emotionellt från 0 till 100 på ingen tid alls, för att sedan kattmjukt bromsa in och parkera, lugnt och behärskat. Och Robert är tämligen oemotståndlig. Här har Harold åstadkommit ännu en intelligent komedi för alla oss vänner av intelligens och komedi, gärna i kombination. SvD We found that if one summarizes the text to 30 percent of original length one will obtain around percent accuracy on 3-4 pages news articles... but query based evaluations are based on subjective opinions These evaluation need large human effort Small overlap of opinions We need man-made extracts to compare the machine made extracts automatically

6 31 32 The yellow sections are the human extracts. The union is SweSum selection Automatically recognize person names, organizations, geographical locations, and time points. Language dependent Person names: Erik Ericsson, Dr. Ericsson Geographical locations: Stockholm, LA, Getaryd, Helsingborg,Valhallavägen Organizations: SBAB, Ericsson AB, SJ, KTH, Statskontoret, Pressbyrån Time points: Torsdagen, 4 maj 2004, 20:00, eftermiddagen. index.html Both rules and training Dictionary and matching rules Training corpora containing Swedish news texts

7 Person names: first names and last names with dictionary Hans Sylvander, Hans van der Vriees Herr Åström, Mr Erik Åström, Ms, Mrs Dr. Åström, Prof Åström Vice VD Erik Åström Erik Karl Johan Åström Places: Geographical dictionary -vägen väg, -gatan, -gata, -området -parken, -park, -land, -rike, -stad, -ryd, -bo, -borg, -by Organizations : Dictionary with companies More than two capital letters following each other e.g AU-systems, KLM, -fabriken, -verk, -företaget, -firma, -byrå AB Spik och Plank Time points: Dictionary with days and months januari, måndag, tisdag, måndag, torsdag, tor, tors den 10:e januari 2001, 10:e januari, 10 januari, 2001 kortformer , 1/ , 1/10 10:14,

8 New technology Based on Random indexing Train the summarizer on a large amount of text Find co-occurencies of terms Let HolSum propose good summaries, compare summaries and found co-occurencies. Choose the optimal summary. Co-occurencies are good. Result as good as traditional approaches Columbia Newsblaster Many news articles summarized into one article There exist many algorithms for this but one can use SweSum text summarization system as a base. 1) Summarize all texts separately 2) Take all summaries and summarize them to one text. But some texts might not be relevant, we must remove them. Cluster text Each cluster should contain only relevant texts ) Cluster say 100 texts in say 10 clusters. (Each cluster can produce one multi-text summary) 2) Rank each text in each cluster 3) Summarize each text and compile, (concatenate) each text together into one larger text. 4) Summarize each large text into one multi text summary => We obtain 10 multi-text summaries. Tagging instead of static lexicons Clause level summarization Improved Pronominal Resolution Automatic evaluation method

9 SweSum Standard version SweSum Experimental NE version swesum_lab/index-eng.html (SweSum uses a Perl-CGI script, there is also a standalone version for plain text/html) Automatic text summarization will be widely used Business Intelligence News publishing News paper, web, WAP, 3G, TV, Radio Machine translation preprocessing Advanced Search engines, Keyword and Named Entities extraction Semantic Webb Human language technology will play key role