Upload
savannah-mcguire
View
29
Download
0
Embed Size (px)
DESCRIPTION
COGEX at the Second RTE. Marta Tatu, Brandon Iles, John Slavick, Adrian Novischi, Dan Moldovan Language Computer Corporation April 10 th , 2006. LCC’s Submission to RTE2. Linear combination of three entailment scores COGEX with constituency parse tree-derived logic forms - PowerPoint PPT Presentation
Citation preview
COGEX at the Second RTE
Marta Tatu, Brandon Iles, John Slavick, Adrian Novischi, Dan Moldovan
Language Computer CorporationApril 10th, 2006
COGEX@RTE2 2
LCC’s Submission to RTE2
Linear combination of three entailment scores
1. COGEX with constituency parse tree-derived logic forms2. COGEX with dependency parse tree-derived logic forms3. Lexical alignment between T and H
For each pair i (Ti,Hi)
If then Ti entails Hi
Lambda (λ) parameters learned on the development data for each task (IE, IR, QA, SUM)
5.0)()()( iscoreλiscoreλiscore LexAlignLexAlignCOGEXCOGEXCOGEXCOGEX DDCC
COGEX@RTE2 3
Approach to RTE with COGEX
Transform the two text fragments into 3-layered logic forms Syntactic Semantic Temporal
Automatically create axioms to be used during the proof Lexical Chains axioms World Knowledge axioms Linguistic transformation axioms
Load COGEX’s SOS with T and H and its USABLE list of clauses with the generated axioms,
Search for a proof by iteratively removing clauses from SOS and searching the USABLE for possible inferences until a refutation is found If no contradiction is detected
Relax arguments Drop entire predicates from H
Compute proof score
semantic and temporal axioms
COGEX@RTE2 4
COGEX Enhancements (1/3)
Logic Form Transformation
Negations not_RB(x1,e1) & walk_VB(e1,x2,x3) » -
walk_VB(e1,x2,x3)
not_RB(x1,e1) & walk_VB(e1,x2,x3) & fast_RB(x4,e1) » -fast_RB(x4,e1)
no/DT case_NN(x1) & confirm_VB(e1,x2,x1) » -confirm_VB(e1,x2,x1)
COGEX@RTE2 5
COGEX Enhancements (1/3)
Logic Form Transformation Temporal normalization of date/time predicates
13th of January 1990 vs. January 13th, 1990
13th_of_January_1990_NN(x1) vs. January_13th_1990_NN(x1)
time_TMP(BeginFN(x1), year, month, day, hour, minute, second) & time_TMP(EndFN(x1), year, month, day, hour, minute, second)
time_TMP(BeginFN(x1), 1990, 1, 13, 0, 0, 0) & time_TMP(EndFN(x1), 1990, 1, 13, 23, 59, 59)
COGEX@RTE2 6
COGEX Enhancements (1/3)
Logic Form Transformation
Temporal context SUMO predicates (Clark et al., 2005)
(S,E1,E2) : S is the temporal signal linking two events E1 and E2
during_TMP(e1,x1), earlier_TMP(e1,x1), …
COGEX@RTE2 7
Logic Forms Differences
Generate LF from two different sources Constituency parse of the data Dependency parse trees (data provided by the
challenge organizers)
Constituency DependencySemantic informationTemporal information
Captures better the (long-range) syntactic dependenciesTemporal normalization (only)NEs imported from the constituency LF whenever the tokens matched (no control over tokenization)
COGEX@RTE2 8
Logic Forms Differences
Gilda Flores was kidnapped on the 13th of January 1990.
Constituency: Gilda_NN(x1) & Flores_NN(x2) & nn_NNC(x3,x1,x2) & _human_NE(x3) & kidnap_VB(e1,x9,x3) & on_IN(e1,x8) & 13th_NN(x4) & of_NN(x5) & January_NN(x6) & 1990_NN(x7) & nn_ NNC(x8,x4,x5,x6,x7) & _date_NE(x8) & THM_SR(x3,e1) & TMP_SR(x8,e1) & time_TMP(BeginFN(x1), 1990, 1, 13, 0, 0, 0) & time_TMP(EndFN(x1), 1990, 1, 13, 23, 59, 59) & during_TMP(e1,x8)
Dependency: Gilda_Flores_NN(x2) & _human_NE(x2) & kidnap_VB(e1,x4,x2) & on_IN(e1,x3) & 13th_NN(x3) & of_IN(x3,x1) & January_1990_NN(x1)
COGEX@RTE2 9
COGEX Enhancements (2/3)
Axioms on Demand Lexical Chains
Consider the first k=3 senses for each word
Maximum length of a lexical chain = 3
DERIVATIONAL WordNet relation is ambiguous with respect to the role of the noun
Derivation-ACT: employ_VB(e1,x1,x2) → employment_NN(e1)
Derivation-AGENT: employ_VB(e1,x1,x2) → employer_NN(x1)
Derivation-THEME: employ_VB(e1,x1,x2) → employee_NN(x2)
Morphological derivations between adjectives and verbs
COGEX@RTE2 10
COGEX Enhancements (2/3)
Axioms on Demand Lexical Chains
Augment with the NE predicate for NE target concepts nicaraguan_JJ(x1,x2) → Nicaragua_NN(x1) &
_country_NE(x1) Discard lexical chains
with more than 2 HYPONYMY relations (H too specific) with a HYPONYMY followed by an ISA
Chicago_NN(x1) → Detroit_NN(x1) which include general concepts: object/NN, act/VB, be/VB
ni = number of hyponyms of concept ci
N = number of concepts in ci’s hierarchy
)1log(
)1log()(
N
ncWgenerality i
i
COGEX@RTE2 11
More Axioms
Another 73 World Knowledge axioms
Semantic Calculus – combinations of two semantic relations (82 axioms) ISA, KINSHIP, CAUSE are transitive relations
ISA_SR(x1,x2) & PAH_SR(x3,x2) → PAH_SR(x3,x2) Mike is a rich man → Mike is rich
Temporal Reasoning Axioms (Clark et al., 2005) (65 axioms) Dates entail more general times
October 2000 → year 2000
during_TMP(e1,e2) & during_TMP(e2,e3) → during_TMP(e1,e3)
COGEX@RTE2 12
COGEX Enhancements (3/3)
Proof Re-Scoring
(T) smart people → people (H)
(T) people → smart people (H)
Entities mentioned in T and H are existentially quantified
Universally quantified T and H entities
(T) people → smart people (H)
(T) smart people → people (H)
COGEX@RTE2 13
Shallow Lexical Alignment
Compute the edit distance between T and H Cost (deletion of a word from T) = 0
Cost (replace of a word from T with another in H) = ∞
Cost (insert a word from H) =
Edit distance between synonyms = 0
T:The Council of
Europehas
45 member states.
Three countries from …
DEL INS DEL
H:The Council of
Europeis made up
by45 member
states.
otherwise 10
verbsWNfor 13
adv and adj nouns, WNfor 30
,
,
,
COGEX@RTE2 14
Results
Learned parameters: IE: score given by COGEXC with some correction from COGEXD
IR: the highest contribution is made by LexAlign (~62%)
COGEXD better on IE, IR, QA (~69% accuracy)
COGEXC better on SUM (~66% accuracy)
Three-way combination outperforms any individual results and any two-system combination
COGEX@RTE2 15
Results, Future Work
Higher accuracy on the SUM task SUM is the highest accuracy task for all systems
(false entailment pairs had H completely unrelated with the texts T)
IE: highest number of false positives Future enhancements
Other types of context: report, planning, etc. Need for more axioms
Automatic gathering of semantic axioms
Paraphrase acquisition (phrase1 → phrase2)
Thank You !
Questions?