The MUC conferences

• How do the MUC conferences measure the performance of each text understanding system?
• What is recall?
• What is precision?
• What's an easy way to get 100% recall?
• What's a (relatively) easy way to approach 100% precision?
• True or False: Good performance on precision and recall at the MUCs was right around 50% for recall and 60% for precision.

Dealing with unknown words

• The first thing TACITUS does is deal with _________ words.
• What tactic did TACITUS use to identify unknown words as Spanish surnames?

Filtering out irrelevant sentences.

• How does TACITUS simplify the task of understanding whole messages, whose sentence length averages 25-30 words?
• How does TACITUS identify relevant sentences in a message?

Template Activation

• In the text understanding system described by Jacobs and Rao, what activates templates?
• What happens when a template gets activated?
• What is a dynamic lexicon in the sense used by Jacob and Rao? Why not just put all your lexical knowledge in the core lexicon?
• What is an example of a word which might have a domain-specific usage, which would best be defined as part of the dynamic lexicon rather than the core lexicon?

Preprocessing

• Text understanding techniques often use a preprocessing phase which uses _______________ techniques to _____________ the sentence into phrases.

Pre-processing

• What is the purpose of the pre-processing phase in automated text understanding?

Parsing

• What is the output of the parsing module of TACITUS?
• What is the main disadvantage of bottom-up parsing?
• What is the main advantage of bottom-up parsing?
• How does TACITUS prune the parsing space?
• How does TACITUS deal with very long sentences?

Analysis

• In Jacobs and Rao's text understanding system, what does the parser provide to the relation-driven control module? What does it get back from it?
• In Jacobs and Rao's text understanding system, semantic interpretation is done largely by _______ing the _____________ for the events and sub-events which pertain to active schema templates.

Semantic interpretation

• Say the parser produces these propositions: 'there exists x, and x is a train, and there exists y, and y is dynamite, and x hit y and x derailed'. What is the job of the semantic interpreter in TACITUS?
• What kind of knowledge does the semantic interpreter refer to in order to fill in the semantic gaps in text?
• When the semantic interpreter has two possible interpretations, how does it choose between them?

Relation-driven control

• In the illustration, what does the HEAD indicate?
• In the illustration, what does the ROLE indicate?
• In the illustration, what does the FILLER indicate?
• In the illustration what kind of question is being asked along the edge marked 'A'?
• In the illustration what kind of question is being asked along the edge marked 'B'?
• In the illustration what kind of question is being asked along the edge marked 'C'?
• Where does the relation-driven control module get inputs that the illustration describes?
• What does the relation-driven control module do with the inputs that the illustration describes?

Machine Understanding and Data Extraction

• The input to a data extraction program is a stream of text, possibly from newswires. In the case of the data extraction exemplfied in the MUC conferences, what is the output?