본문 바로가기
KERD
논문KCI 등재

Language model performance on English control constructions and its implications

이수환
언어과학회 (The Linguistic Science Society)|2024.06|언어과학연구 (The Journal of Linguistics Science)|제109호|pp. 245-278|DOI 10.21296/jls.2024.06.109.245

2

조회

0

다운로드

0

인용

0

참고문헌

키워드

language acquisitionparameterslanguage modelscontrol constructions
주제분류언어학

초록

This paper examines the learnability of subject and object control constructions with transitive verbs (e.g., promise and persuade). It explores how children and Transformer-based language models (LMs), such as GPT-2, distinguish these constructions using semantic, contextual, and syntactic cues. While children rely on syntactic cues such as double object constructions (DOCs) to disambiguate subject and object control verbs (Becker, 2014), GPT-2 derives a different conclusion with respect to this distinction. The study highlights the importance of accessing the underlying syntactic structures of control predicates in helping children distinguish between these constructions (Larson, 1991). It also discusses implications for improving LM performance, suggesting that allowing LMs the access to underlying structures during training could enhance their accuracy in handling surface-ambiguous structures.

Abstract

This paper examines the learnability of subject and object control constructions with transitive verbs (e.g., promise and persuade). It explores how children and Transformer-based language models (LMs), such as GPT-2, distinguish these constructions using semantic, contextual, and syntactic cues. While children rely on syntactic cues such as double object constructions (DOCs) to disambiguate subject and object control verbs (Becker, 2014), GPT-2 derives a different conclusion with respect to this distinction. The study highlights the importance of accessing the underlying syntactic structures of control predicates in helping children distinguish between these constructions (Larson, 1991). It also discusses implications for improving LM performance, suggesting that allowing LMs the access to underlying structures during training could enhance their accuracy in handling surface-ambiguous structures.

참고문헌 (62)

  1. Arunachalam, S., "Preschoolers’ acquisition of novel verbs in the double object dative", Cognitive Science, 41, pp.831~854 (2017)
  2. Becker, M., "The acquisition of syntactic structure: Animacy and thematic alignment, vol. 141" (2014), Cambridge University Press
  3. Belletti, A., "From grammar to meaning: The spontaneous logicality of language", pp.294~ (2013), Cambridge University Press
  4. Bender, E. M., "Climbing towards NLU: On meaning, form, and understanding in the age of data", Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp.5185~5198 (2020)
  5. Boeckx, C., "Reply to “control is not movement”", Linguistic Inquiry, 34(2), pp.269~280 (2003)
  6. Boeckx, C., "Icelandic control really is A-movement : Reply to Bobaljik and Landau", Linguistic Inquiry, 41(1), pp.111~130 (2010)
  7. Bruening, B., "Qr obeys superiority : Frozen scope and ACD", Linguistic Inquiry, 32(2), pp.233~273 (2001)
  8. Bruening, B., "Double object constructions disguised as prepositional datives", Linguistic Inquiry, 41(2), pp.287~305 (2010)
  9. Bruening, B., "Implicit arguments in English double object constructions", Natural Language & Linguistic Theory, 39(4), pp.1023~1085 (2021)
  10. Campbell, A. L., "The acquisition of English dative constructions", Applied Psycholinguistics, 22(2), pp.253~267 (2001)
  11. Chomsky, C., "5 to 10" (1969), The MIT Press
  12. Collins, C., "A smuggling approach to raising in English", Linguistic Inquiry, 36(2), pp.289~298 (2005)
  13. Collins, C., "A smuggling approach to the passive in English", Syntax, 8(2), pp.81~120 (2005)
  14. "Summary: Subject control verb PROMISE in English" (1998)[URL]
  15. Culicover, P. W., "Control is not movement", Linguistic Inquiry, 32(3), pp.493~512 (2001)
  16. Dai, Z., "Transformer-XL : Attentive language models beyond a fixed-length context", Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp.2978~2988 (2019)
  17. "The corpus of contemporary American English: 450 million words, 1990-present" (2008)[URL]
  19. Gleitman, L., "The structural sources of verb meanings", Language Acquisition, 1(1), pp.3~55 (1990)
  20. "Assessing BERT’s syntactic abilities" (2019)[URL]
  21. Goodluck, H., "Studies in the acquisition of anaphora", pp.247~269 (1987), Springer
  22. Goodluck, H., "The nominal analysis of children’s interpretations of adjunct pro clauses", Language, 77, pp.494~509 (2001)
  23. Grimshaw, J., "Argument structure" (1990), The MIT Press
  24. Hornstein, N., "Movement theory of control", pp.1~41 (2010), John Benjamins
  25. Hovav, M. R., "The English dative alternation : The case for verb sensitivity", Journal of Linguistics, 44(1), pp.129~167 (2008)
  26. Hsu, J. R., "Control and coreference in early child language", Journal of Child Language, 16(3), pp.599~622 (1989)
  27. Hsu, J. R., "The development of grammars underlying children’s interpretation of complex sentences", Cognition, 20(1), pp.25~48 (1985)
  28. Hu, J., "A closer look at the performance of neural language models on reflexive anaphor licensing", Proceedings of the Society for Computation in Linguistics 2020, pp.274~283 (2020)
  29. Hu, J., "Prompting is not a substitute for probability measurements in large language models", Proceedings of the 2023Conference on Empirical Methods in Natural Language Processing, pp.5040~5060 (2023)
  30. Jackendoff, R., "The semantic basis of control in English", Language, 79, pp.517~556 (2003)
  31. Kirby, S., "Semantic scaffolding in first language acquisition: The acquisition of raising-to-object and object control", Doctor, pp.2009
  32. Kirby, S., "Up to date on raising and control part 1 : Properties and analyses of the constructions", Language and Linguistics Compass, 4(6), pp.390~400 (2010)
  33. Landau, B., "Language and experience: Evidence from the blind child" (1985), Harvard University Press
  34. Landau, I., "The explicit syntax of implicit arguments", Linguistic Inquiry, 41(3), pp.357~388 (2010)
  35. Landau, I., "A selectional theory of adjunct control" (2021), MIT Press
  36. Larson, R. K., "On the double object construction", Linguistic Inquiry, 19(3), pp.335~391 (1988)
  37. Larson, R. K., "Promise and the theory of control", Linguistic Inquiry, 22(1), pp.103~139 (1991)
  38. Lee, S. H., "Can language models capture syntactic associations without surface cues? A case study of reflexive anaphor licensing in English control constructions", Proceedings of the Society for Computation in Linguistics 2022, pp.206~211 (2022)
  39. Levin, B., "English verb classes and alternations: A preliminary investigation" (1993), University of Chicago Press
  40. Linzen, T., "How can we accelerate progress towards human-like linguistic generalization?", Proceedings of the 58th annual meeting of the Association for Computational Linguistics, pp.5210~5217 (2020)
  41. Linzen, T., "Assessing the ability of LSTMs to learn syntax-sensitive dependencies", Transactions of the Association for Computational Linguistics, 4, pp.521~535 (2016)
  42. MacWhinney, B., "The Childes Project: Tools for analyzing talk, vol. 2: The database" (2014), Psychology Press
  43. Marvin, R., "Targeted syntactic evaluation of language models", Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp.1192~1202 (2018)
  44. Mateu, V., "Smuggling in syntax", pp.222~254 (2021), Oxford University Press
  45. McDaniel, D., "Language processing and language acquisition", pp.313~325 (1990), Springer
  47. Müller, G., "Incomplete category fronting", pp.1~28 (1998), Springer
  48. Orfitelli, R., "‘Experiencing’ a slight delay: Intervening arguments and the acquisition of subject-to-subject raising", Proceedings of the 36th Boston University Conference on Language Development, 2, pp.374~386 (2012)
  49. Pinker, S., "Language learnability and language" (1984), Harvard University Press
  50. Pylkkänen, L., "Introducing arguments, vol. 49" (2008), MIT Press
  51. "Language models are unsupervised multitask learners" (2019)[URL]
  52. Rizzi, L., "Relativized minimality" (1990), The MIT Press
  53. Rosenbaum, P. S., "The grammar of English predicate complement constructions", pp.1965
  54. Rosenbaum, P. S., "Phrase structure principles of English complex sentence formation", Journal of Linguistics, 3(1), pp.103~118 (1967)
  55. Sennrich, R., "Neural machine translation of rare words with subword units", Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pp.1715~1725 (2016)
  56. Snyder, W., "The structure and acquisition of English dative constructions", Linguistic Inquiry, pp.281~317 (1997)
  57. Vaswani, A., "Attention is all you need", Advances in neural information processing systems, pp.5998~6008 (2017)
  58. Warstadt, A., "BLiMP : The Benchmark of Linguistic Minimal Pairs for English", Transactions of the Association for Computational Linguistics, 8, pp.377~392 (2020)
  59. Wilcox, E., "What do RNN language models learn about filler–gap dependencies?", Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pp.211~221 (2018)
  60. Wolf, T., "Transformers : State-of-the-art natural language processing", Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System demonstrations, pp.38~45 (2020)
  61. Yang, Charles, "The price of linguistic productivity: How children learn to break the rules of language" (2016), MIT press
  62. Zhang, Y., "When do you need billions of words of pretraining data?", Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pp.1112~1125 (2021)

관련 논문AI추천

How Language Models Process Subject and Object Control: On Promise as a Subject Control Verb노강산한국생성문법학회 · 생성문법연구2024 · 제34권 제4호Syntax and Statistical Learning in the Language Development of the English Dative Alternation김유희, 박명관대한영어영문학회 · 영어영문학연구2013 · 제55권 제4호BERT Learns More than Word Frequency Information: A Case Study of Do-Be Constructions송상헌, 신운섭한국언어학회 · 언어2022 · 제47권 제3호An acquisition-based evaluation metric for competing syntactic theoriesIsaac Gould한국현대언어학회 · 언어연구2018 · 제35권 제1호Syntactic priming in the L2 neural language model박명관, 최선주언어과학회 · 언어과학연구2022 · 제103호언어 학습자로서의 신경망 언어모델: 영어 주어-동사 일치를 중심으로김유희, 박명관, 정원일언어정보연구소 · 언어와 정보 사회2021 · 제44권

이 간행물의 다른 논문 (10건)

A corpus-based analysis of English vowel production by Korean L2 learners유혜배pp. 1-24언어접촉 시각에서 본 좡어(壯語) 구차용어 ‘tsaŋ2’의 기능 변화: 징시 좡어(靖西壯語)를 중심으로이연정pp. 25-48Syllable dependent loanword adaptation from English to Korean and Chinese김종미, 박미선, 호취정pp. 49-78한국어 청자의 영어 자음 지각 양상 연구남영자pp. 79-95토익 독해 문항 유형 및 언어 능숙도에 따른 한국 대학생들의 전략 사용에 관한 연구지희정pp. 97-122
원문 보기