Automated word puzzle generation using topic models and semantic relatedness measures

Transcription

1 Automated word puzzle generation using topic models and semantic relatedness measures Balázs Pintér, Gyula Vörös, Zoltán Szabó and András Lőrincz ELTE IK

2 Table of contents 1 Introduction Our goal The method 2 3 The performance of the three topic models Some interesting puzzles

3 Our goal Introduction Our goal The method Word puzzles Are used in education, psychometry, etc. (e.g., TOEFL) Are costly to design and maintain Our goal is to generate word puzzles from unstructured and unannotated corpora Puzzle types Odd one out: salmon, shark, whale, elephant Choose the related word: regiment, battalion, army infantry, service, king Separate the topics: water, heat, temperature, pressure superman, clark, luthor, kryptonite

4 The method Introduction Our goal The method Building blocks of puzzles Consistent sets (sets of related words): {salmon, shark, whale} Less related words: elephant Model the corpus as a combination of latent topics Identify consistent sets from among these topics Generate the puzzles by mixing these sets with less related elements Building blocks of the algorithm Topic models Semantic similarity measures Network flow

5 Topic models Introduction Topics Weights Document govern, 0.6 presid, 0.5 forc, 0.2 languag, 0.5 linguist, 0.5 group, 0.2 The government introduced a new voice controlled system that allows air force pilots to control aircraft by issuing commands in natural language... vowel, 0.7 voic, 0.6 conson, 0.6 aircraft, 0.4 air, 0.3 flight, 0.3 standard, 0.3 system, 0.1 implement, 0.1

6 Topic models used Introduction Latent Semantic Analysis arg min X ^X = ^U^S^V T. (1) rank(ˆx)=d F Online Group-Structured Dictionary Learning 1 M ( ) i ρ [ ] 1 min D,{α i } M Mj=1 (j/m) ρ M 2 x i Dα i κω(α i ) (κ > 0), i=1 i=1 (2) 1 Ω(α) = η η α Gj, 2 j (3) Latent Dirichlet Allocation N K M j P(W, Z,θ,φ α,β) = P(φ i β) P(θ j α) P(z j,t θ j )P(w j,t φ zj,t ), (4) i=1 j=1 t=1

7 Topic models used Introduction Latent Semantic Analysis arg min X ^X = ^U^S^V T. (1) rank(ˆx)=d F Online Group-Structured Dictionary Learning 1 M ( ) i ρ [ ] 1 min D,{α i } M Mj=1 (j/m) ρ M 2 x i Dα i κω(α i ) (κ > 0), i=1 i=1 (2) 1 Ω(α) = η η α Gj, 2 j (3) Latent Dirichlet Allocation N K M j P(W, Z,θ,φ α,β) = P(φ i β) P(θ j α) P(z j,t θ j )P(w j,t φ zj,t ), (4) i=1 j=1 t=1

8 health jump treatment medical class fence care 0.29 (a) A consistent set patients rider 0.15 horse (b) An inconsistent set

9 Odd one out Mix a consistent set and a less related word salmon, shark, whale, elephant Choose the related word Mix a consistent set with some less related words Present the words in a different grouping regiment, battalion, army infantry, service, king Separate the topics Mix two (or more) consistent sets water, heat, temperature, pressure superman, clark, luthor, kryptonite

10 The performance of the three topic models Some interesting puzzles The performance of the three topic models number of consistent sets OSDL LDA LSA number of consistent sets OSDL LDA LSA threshold (c) Wikipedia threshold (d) NIPS proceedings

11 Some interesting odd one out puzzles The performance of the three topic models Some interesting puzzles Consistent set of words Odd one out cao wei liu emperor king superman clark luthor kryptonite batman devil demon hell soul body egypt egyptian alexandria pharaoh bishop singh guru sikh saini delhi language dialect linguistic spoken sound mass force motion velocity orbit voice speech hearing sound view athens athenian pericles corinth ancient data file format compression image function problems polynomial equation physical