>>> Identifying Relationships in Unstructured Texts (Rani Nelken, Elif Yamangil, Wang Xiaoxuan)

Created on : July 29, 2015

From Named Entity Recognition to Identifying Relationships in Unstructured Texts

Title slide: Identifying Relationships in Unstructured Texts, by Elif Yamangil and Rani Nelken, Harvard School of Engineering and Applied Sciences.

Slide 'Goal: From Text to DB': an arrow from a pile of text fragments to a database cylinder; the text holds the information, the database can answer questions.

Slide 'Example: Chinese Biographies': aggregate questions such as average lifespan, filtering by date or geography, and identifying classes of people.

Slide with the single statement: the current manual approach equals a bottleneck.

Slide 'Solution: Automatic Text Understanding', answered by the HAL 9000 line, I'm sorry Dave, I'm afraid I can't do that.

Slide with the single statement: language is ambiguous.

Slide 'What's practical?', answered with two words: Information Extraction.

Slide 'Information Extraction': extract the pieces salient to the user's needs - who did what to whom - and link back to the original text for reference.

Slide 'Applications: Business', with the single example mergers and acquisitions.

Slide 'Applications: Biomedical', with the single example mining the bibliome.

Slide on filling database slots from sub-segments of text: a 2002 news story about Microsoft becomes a table of name, title and organization.

Slide 'Main Challenges' beside a graph of red nodes and arrows: who did what to whom, named entity resolution for who and whom, relation extraction for what.

Slide showing the same Microsoft news story with three extracted triples - Microsoft, CEO, Bill Gates and so on - grouped by braces into a database cylinder.

Slide 'Named Entity Recognition': find names of people, locations and offices, classify types, link variants such as Mr. Gates and William G., resolve anaphora.

Slide 'How we do it?': Searle's Chinese Room thought experiment, with a cartoon of a person on a chair passing a note under a closed door.

Slide 'Incremental approach': step 1, manually defined patterns; step 2, supervised learning.

Slide 'Manual patterns': appropriate when specific information is needed, such as converting a date to AD; built incrementally as a pipeline of patterns.

Slide 'Manual patterns: howto': domain experts specify the patterns, they are encoded as regular expressions, run on the database, analysed, then refined.