FACT: A Learning Based Web Query Processing System Hongjun Lu, Yanlei Diao Hong Kong U. of Science & Technology Songting Chen, Zengping Tian Fudan University

Demonstration, SIGMOD Outline  Introduction  Learning Based Web Query Processing  FACT: A Prototype System  Preliminary System Evaluation  Conclusions

Demonstration, SIGMOD How Do We Query the Web?  Use a search engine  Form query key words  An example: Find room rates of hotels in Hong Kong  used search engine  keywords: Hong Kong+hotel

Demonstration, SIGMOD Hotel 2 Hotel 1 3 done forward Look at the Number!

Demonstration, SIGMOD Query the Web -- Current Situation  Search engines return a long list of URLs. User is required to browse the web pages to find the information.  The information required is often not on the returned page -- navigation through hyperlinks is often required (those links may or may not that obvious).  The target information is in different forms (paragraphs, lists, tables …)  A lot of web pages to be browsed Are we happy with this?

Demonstration, SIGMOD Efforts to Improve the Situation  Search engines  better index, improve precision/recall, metasearch engines, better presentation of results, ….  IR techniques to Web  document clustering/indexing, better model, similarity functions, documents ranking,...  Intelligent agent  user profiling, hyperlink recommendation,...  Database approach  wrappers, query languages, …

Demonstration, SIGMOD Our Dream  Querying the Web as easy as querying a relational database  SQL query returns a table of hotel prices SELECT room rates FROM web.hotel WHERE city = “hong kong”  May remain a dream for a while :-(

Demonstration, SIGMOD A Practical goal  Use keywords to express query requirements  simple, no need to know schema of data  inaccurate  Relieve users from tedious browsing as much as possible  Not URLs, not Web sites, even not Web pages  Present query results to users as accurate and concise as possible  Tables, lists, paragraphs, … containing user required information

Demonstration, SIGMOD Query Results -- Queried Segments  Return query results as accurate and concise as possible.  Basic idea:  Breaking a Web page into segments: a row in a table, a table, an item in a list, a list, a paragraph,  returning only queried segments to users queried segments : segments that contain the information the user is interested in.

Demonstration, SIGMOD Outline  Introduction  Learning Based Web Query Processing  FACT: A Prototype System  Preliminary System Evaluation  Conclusions

Demonstration, SIGMOD Learning Based Query Processing  The fundamental difficulties in Web query processing:  Web is a huge, ever growing, heterogeneous, semi-structured data source  Most users of Web are naïve users issuing ad hoc queries  Learn the knowledge for query processing from the User!

Demonstration, SIGMOD A Learning Based Technique  Learn from the user when he browses from the first few URLs  to navigate through the web pages  to identify the required information in a web page  Process the rest URLs automatically and retrieve queried segments

Demonstration, SIGMOD Hotel 2 Hotel 1 3 done forward User browses it!

Demonstration, SIGMOD Back User clicks here!

Demonstration, SIGMOD Room information User marks it!

Demonstration, SIGMOD back Fact starts here!

Demonstration, SIGMOD roomrates Fact chooses it!

Demonstration, SIGMOD xxx Fact finds it!

Demonstration, SIGMOD Outline  Introduction  Learning Based Web Query Processing  FACT: A Prototype System  Preliminary System Evaluation  Conclusions

Demonstration, SIGMOD A Query Processing System A learning based query processing system:  User Interface: accepts user queries, presents query results, a browser capable of capturing user actions  Query Analyzer: analyzes and transforms user queries  Session Controller: coordinates learning and locating  Learner: generates knowledge from captured user actions  Locator: applies knowledge and locates query results  Crawler & Parser: retrieves pages and parses to trees  Knowledge Base: stores learned knowledge

Demonstration, SIGMOD Reference Architecture Session Controller Locator Search Engine Web User Interface Knowledge Base Learner Query Analyzer Crawler & Parser User

Demonstration, SIGMOD A Query Session Session Controller Training Strategy Segment Graph Result Buffer Knowledge Base User Actions Query results Checking URLs Locating Process Locator Query Result Presenter Learning Process Learner Browser Scripts

Demonstration, SIGMOD Training Strategies  Sequential  First n sites: user browses and system learns  Next N-n sites: system processes  Random  Randomly choose n sites: user browses and system learns  the system processes the rest  Interleaved  First n 0 sites, user browses and system learns  Next n - n 0 site, system makes decision. For incorrect ones, user browses and system re-learns  Next N-n sites: system processes

Demonstration, SIGMOD Outline  Introduction  Learning Based Web Query Processing  FACT: A Prototype System  Preliminary System Evaluation  Conclusions

Demonstration, SIGMOD System Evaluation  Functionality  Performance  precision, recall, correctness  efficiency: in a site, how many pages the system visits to find a result  training efficiency: how many training samples are needed  User interface

Demonstration, SIGMOD

Demonstration, SIGMOD System Evaluation - Effectiveness  Given a set of keywords, the system makes N decisions N =N1 + N2 + N3 + N4 Precision = N1 / (N1+N3), Recall= N1 / # relevant sites, Correctness = (N1+N2) / N.

Demonstration, SIGMOD System Evaluation - Efficiency  How efficiently the system finds a queried segment in a site? Level of a Queried Segment = the length of the shortest path to find it Absolute Path length = # Crawled pages, Relative Path Length = # Crawled pages / Level of the Queried Segment.

Demonstration, SIGMOD Basic Performance Q 11 : Hong Hong Hotel Room Rate Q 12 : Hong Kong Hotel Sequential training

Demonstration, SIGMOD Query Q 12 Effects of training Strategies

Demonstration, SIGMOD Improved Performance Interleaved training

Demonstration, SIGMOD Outline  Introduction  Learning Based Web Query Processing  FACT: A Prototype System  Preliminary System Evaluation  Conclusions

Demonstration, SIGMOD Conclusions  Proposed and implemented learning based Web query processing with the following features  Returning succinct results: segments of pages;  No a prior knowledge or preprocessing, suited for ad hoc queries;  exploiting page formatting and linkage information simultaneously.  The preliminary results are promising

Demonstration, SIGMOD Future Work  Better knowledge  key factor that affects system performance  Dynamic web pages ?  Integrating results from another project  System evaluation  Prototype  product  dot com company $$$ ???