Discovery of Inference Rules for Question Answering Dekang Lin and Patrick Pantel Natural Language Engineering 7(4):343-360, 2001 as (mis-)interpreted by Peter Clark
Goal Observation: “mismatch” between expressions in qns and text e.g. “X writes Y” vs. “X is the author of Y” Need “inference rules” to answer questions “X writes Y” “X is the author of Y” “X manufactures Y” “X’s Y factory” Question: Can we learn these inference rules from text? (aka “paraphrases”, “variants”) DIRT (Discovering Inference Rules from Text)
The limits of word search… Who is the author of ‘Star Spangled Banner?’ A. …Francis Scott Key wrote the “Star Spangled Banner” in 1814. …comedian-acress Roseanne Barr sang her famous shrieking rendition of the “Star Spangled Banner” before a San Diego Padres-Cincinnati Reds game. B. What does Peugot manufacture? Chrétien visited Peugot’s newly renovated car factory in the afternoon.
Approach Parse sentences in a giant (1GB) corpus Extract instantiated “paths” from the parse tree, e.g.: X buys something from Y X manufactures Y X’s Y factory For each path, collect the sets of X’s and Y’s For a given path (pattern), find other paths where the X’s and Y’s are pretty similar
Approach Parse sentences in a giant (1GB) corpus, then: Extract “paths” from the parse tree, e.g.: X buys something from Y X manufactures Y X’s Y factory Collect statistics on what the X’s and Y’s are Compare the X-Y sets: For a given path (pattern), find other paths where the X’s and Y’s are similar
Results (examples)
Method: 1. Parse Corpus 1GB newspaper (Reuters?) corpus Use MiniPar Chart parser self-trained statistical ranking of parse (“dependency”) trees
Method: 2. Extract “paths”
Method: 3. Collect the X’s and Y’s
Method: 4. Compare the X-Y sets ?…
Method: 4. Compare the X-Y sets ? …and
Method: 4. Compare the X-Y sets 1. Characterizing a single X-Y set: Count frequencies of words for X (and Y) Weight by ‘saliency’ (slot-X mutual information)
Method: 4. Compare the X-Y sets 2. Comparing two X-Y sets Two paths have high similarity if there are a large number of common features. Mathematically:
Example: Learned Inference rules
Example: vs. Hand-crafted inference rules (by ISI)
Results
Observations Little overlap in manual and automatic rules DIRT performance varies a lot Much better with verb rather than noun roots If less than 2 modifiers, no paths found For some TREC examples, no “correct” rules found “X leaves Y” “X flees Y” Where X’s and Y’s are similar, can get agent-patient the wrong way round E.g. “X asks Y” vs. “Y asks X”
The Big Question Can we acquire the vast amount of common-sense knowledge from text? Lin and Pantel suggests: “yes” (at least in a semi-automated way)