Lecture 14: lexical semantics Professor Robert C. Berwick [email protected] 6.863J/9.611J SP11 The Menu Bar • Administrivia • What knowledge do we need beyond syntax? (Knowledge about words) • How can we learn this stuff? 6.863J/9.611J SP11 “I saw one car flattened down to about one foot high, and my mechanics friend told me that the driver who got out of that cab that was squashed down by accident got out by a narrow escape” People can learn a lot of language from just a little input 6.863J/9.611J SP11 TADOMA The unlikely channel6.863J/9.611J SP11 The question - lab 5–6 • What lexical-semantic generalizations has the parser made from being trained on the Penn Tree Banks? • Are these the right generalizations? • If not, how can we fix this? 6.863J/9.611J SP11 Verb knowledge Knowledge about ‘alternations’: The moon rotates around the earth The moon revolves around the earth The moon rotates the earth ?The moon revolves the earth John smeared paint on the wall John smeared the wall with paint John sprayed the wall with pain John sprayed paint at the wall ?John smeared paint at the wall 6.863J/9.611J SP11 6.863J/9.611J SP116.863J/9.611J SP11 Lexical semantics task 6.863J/9.611J SP11 Testing the parser • Look at verbs in an alternation class (which is ‘semantically’ and syntactically coherent • Find the –logprobs for the alternations, including the ‘ungrammatical’ ones • Do these match up with intuitions and expectations from frequencies in the Penn Tree Bank? 6.863J/9.611J SP11 ‘Give’ NP NP vs. NP PP-DTV • 256 total give NP NP or NP PP-DTV in PTB • 205 are NP NP 80% • 51 are NP PP-DTV 20% • Which frame is therefore going to be preferred? 6.863J/9.611J SP11 ‘give NP to PP’, sentence 8246.863J/9.611J SP11 6.863J/9.611J SP11 6.863J/9.611J SP11 6.863J/9.611J SP116.863J/9.611J SP11 6.863J/9.611J SP11 6.863J/9.611J SP11 The verb ‘join’ 6.863J/9.611J SP116.863J/9.611J SP11 Caveats • Some verbs have multiple senses • Not all instances of a category label hold the same ‘semantic’ or ‘thematic’ role • To be completely accurate, we’d have to review each and every tree and label each node with a semantic role, very carefully • But as a first approximation, let’s conflate category labels 6.863J/9.611J SP11 6.863J/9.611J SP11 Delete ‘TMP’ nodes • Why? • Temporal PPs, etc. 6.863J/9.611J SP116.863J/9.611J SP11 6.863J/9.611J SP11 6.863J/9.611J SP11 6.863J/9.611J SP116.863J/9.611J SP11 6.863J/9.611J SP11 6.863J/9.611J SP11 A case study: join, merge • “Bristol-Meyers agreed to merge with Sun Microsystems” • “Boeing and Sun Microsystems agreed to merge” • Which would be more likely? Which is more likely? Which ‘should be’ more likely (according to linguistic accounts) 6.863J/9.611J SP11 Some counts • join - 49 VB • mix - 1 • water 114 NN • 24 milk NN • 14 toys NN • 207 computers NNS6.863J/9.611J SP11 Some sentences • John NNP mixed VBD the DT water NN and CC the DT milk NN • John NNP mixed VBD the DT milk NN and CC the DT water NN • John NNP mixed VBD the DT water NN with IN the DT milk NN • John NNP mixed VBD the DT milk NN with IN the DT water NN • John NNP joined VBD the DT water NN and CC the DT milk NN • John NNP joined VBD the DT milk NN and CC the DT water NN • John NNP joined VBD the DT water NN with IN the DT milk NN • John NNP joined VBD the DT milk NN with IN the DT water NN • John NNP joined VBD the DT water NN and CC the DT water NN • John NNP joined VBD the DT water NN with IN the DT water NN • John NNP joined VBD the DT computers NNS and CC the DT computers NNS • John NNP joined VBD the DT computer NNS with IN the DT computer NNS 6.863J/9.611J SP11 The envelope please… • J. mixed the water and the milk • J. mixed the milk and the water • J. mixed the water with the milk • J. mixed the milk with the water • J. joined the water and the milk • J. joined the water with the milk • J. joined the milk with the water • J. joined the water with the milk –log prob: (closer to 0 = more likely) –55.6292 –54.307 –54.3957 –51.2094 –48.4139 –46.1579 –46.1015 –43.0599 6.863J/9.611J SP11 And more • John joined the computers and the computers • John joined the computers with the computers • John joined the milk and the milk • John joined the milk with the milk –39.699 –43.054 –48.0987 –46.3324 6.863J/9.611J SP11 First of all… • John mixed the water with the milk6.863J/9.611J SP11 Then • John mixed the milk with the water Hmm… what about ‘mixed’? Try ‘join’ 6.863J/9.611J SP11 ‘Join’ • John joined the water and the milk 6.863J/9.611J SP11 ‘Join’ • John joined the milk with the water 6.863J/9.611J SP11 In fact… • No matter what lexical item we choose, ‘milk’ (but not ‘water’ or ‘toys’ or ‘computer’) forces a low attachment like this – all the others, in all other combinations, force the high PP attachment…6.863J/9.611J SP11 Where do the numbers come from? A breakdown • John joined the water and the milk • John joined the milk and the water -40.783 -31.8808 -21.5856 -0.6976 -2.202 -2.135 -0.6976 -38.527 -29.6248 -18.715 -2.594 -2.203 6.863J/9.611J SP11 Where do the numbers come from? • John joined the water with the milk • John joined the milk with the water -46.10 -43.06 -29.57 -26.53 -2.19 -15.62 -9.70 6.863J/9.611J SP11 24 ‘milk’ sentences, only a few as a noun… #21219 #5212 6.863J/9.611J SP11 #212736.863J/9.611J SP11 #23482, 23488 #21382, 21314 6.863J/9.611J SP11 6.863J/9.611J SP11 Another example: ‘bark’ 6.863J/9.611J SP116.863J/9.611J SP11 Knowledge of language - not just syntax • Semantic Inference and Language • A step outside syntax • let’s look at computation using • WordNet • project led by a psychologist • George Miller (Princeton University) • handbuilt network of synonym sets (synsets) with semantic relations connecting them • very popular in computational linguistics 6.863J/9.611J SP11 WordNet • what is it? • synonym set (synset) network for nouns, verbs, adjectives and adverbs • synsets connected by semantic relations (isa, antonymy,...) •
View Full Document
Unlocking...