Pith. sign in

REVIEW 4 major objections 6 minor 104 references

Position: Natural Language Should Not Fully Replace Formal Languages

T0 review · 4 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proves a specificity crossover theorem: once task requirements pass a threshold, natural language cannot be more efficient than formal specification, so natural language will never fully replace formal languages.

desk verdict The qualitative complementarity argument is fine; the formal theorem is not proven, and Proposition 3.8 is false as stated. read the letter →

arxiv 2607.20432 v1 pith:QXGQGTQG submitted 2026-05-10 cs.CL cs.PL

Eitan Wagner , Elisha Rosensweig , Omri Abend This is my paper
classification cs.CLcs.PL
keywords taskspecificityunderspecificationtranslationgapcrossoverformallanguagesnaturallanguagegenerationhybridhuman-AIinterfacescode
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This position paper argues that natural language will not, and should not, fully replace formal languages, and gives the argument a precise information-theoretic form. The authors define task specificity—how many bits of genuine requirement information a task forces the user to communicate about the intended output—and show that natural language's advantage comes from underspecification: it can leave many acceptable outputs undescribed. They prove a specificity crossover theorem: for any task whose specificity reaches a threshold, a natural-language description cannot be shorter than a direct formal specification, and formal notation is necessarily at least as efficient. If this is right, the claim that LLMs will make programming languages obsolete is not just practically doubtful but structurally impossible to achieve through scaling alone. The constructive conclusion is hybrid systems that let users move between loose natural-language intent and precise formal detail.

What carries the argument

The framework's load-bearing objects are task specificity, the translation gap, and formal-language redundancy, combined in the specificity crossover theorem. Task specificity is the minimal mutual information that a generated output must share with the intended output for the task to count as satisfied—in effect, the number of bits of genuine requirement information that have to cross the channel. The translation gap is defined as the divergence between the distribution of content in the target domain and the distribution of natural language as a general-purpose code; it is the overhead incurred whenever a language optimized for the whole world is used to describe one narrow domain. Formal-

What would settle it

Find any one domain and one task at or above the predicted crossover threshold where a user's natural-language prompt is shorter than the direct formal specification and still reliably produces an output satisfying every stated constraint, with no extra rounds of iteration. Such a case would violate the lower-bound inequality that the crossover theorem is built on; alternatively, measuring near-zero cost for zero-specificity tasks would directly test the same bound.

Watch

Extended reading notes

Core claim

The central discovery is a proof that there is a crossover threshold in task specification. The paper defines the specificity of a task as the minimum mutual information between the user's intended output and a generated output that satisfies the task; zero specificity means any output is acceptable, while maximal specificity means only the exact intended output will do. It then proves the specificity crossover theorem: with translation gap (the irreducible cost of describing a technical domain in general-purpose natural language) and formal-language redundancy, the threshold is σ* = H(Y) + R_formal − Δ_trans. Below σ*, natural language can be more efficient by exploiting underspecification;

Load-bearing premise

The whole theorem leans on the assumption that achieving a task's required specificity always costs at least that specificity plus the full translation-gap overhead in natural language—even when the task is so loose that any output is acceptable and a one-word prompt would do; if that lower bound fails, the crossover threshold is not established.

Editorial extensions

If this is right

  • In any domain with a reasonably efficient formal language, tasks that require high precision can never be specified more efficiently in natural language, however capable the model becomes.
  • The practical implication is not to abandon natural-language programming but to design hybrid workflows, using natural language where many outputs are acceptable and formal notation where exactness is mandatory.
  • Benchmarks that measure only end-to-end natural-language-to-code generation overstate the value of the pure natural-language path; they should compare natural-language specification, code completion, and unaided coding across a range of task specificities.
  • Observed user behavior in text-to-image and coding tools—prompts growing longer and shifting to model-specific phrasing—is the expected empirical signature of the crossover: users climb the ladder of specification as their requirements tighten.
  • For open-ended, exploratory, or low-expertise tasks, natural language retains a real and theoretically grounded advantage, so it should be the interface of choice there, not a stopgap.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct way to test the framework quantitatively would be to estimate the translation gap and formal-language redundancy for a concrete domain and predict the specificity at which users actually switch from prompting to direct specification; the paper's case studies are qualitative, so this measurement remains open.
  • The crossover threshold is an engineering dial, not a fixed law: reducing the translation gap—by learning model-specific prompt languages, or by building interfaces that already know the domain—shifts the threshold upward, potentially making natural language viable for more tasks.
  • If the lower-bound assumption holds, advanced 'natural language programming' will in the limit resemble a formal language with natural-language syntax: users will have to be precise enough to pay the full translation cost, so the distinction between the two modalities will blur from the user side.
  • The zero-specificity edge case is the sharpest place to probe the bound: the lower bound predicts a positive fixed cost even when any output is acceptable, so measuring near-zero-specificity prompt costs would either confirm the bound or force a refinement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper argues that natural language (NL) and formal languages (FL) are complementary rather than substitutable, and that NL should not fully replace FL. It proposes an information-theoretic framework: tasks are treated as constraint-satisfaction problems; task specificity σ is defined as a minimal mutual information quantity; the translation gap Δtrans is defined as D_KL(pM || pNL); formal-language redundancy is R_formal = L_formal − H(Y). The main formal result is Theorem 3.10, which defines a crossover threshold σ* = H(Y) + R_formal − Δtrans and claims that for σ ≥ σ* natural language cannot be more efficient than formal specification. The paper then applies the framework to image generation, code synthesis, audio, and text, concluding with a call for hybrid systems.

Significance. The manuscript addresses a timely and important question and is clearly written. It provides a valuable synthesis of linguistic concepts (underspecification, pragmatic inference, QUD) and connects them to an information-theoretic framing. The qualitative thesis—that NL is well suited to underspecified, open-ended tasks while formal languages are preferable for high-precision tasks—is plausible and consistent with existing empirical observations, including the cited Midjourney prompting studies. The paper also makes a useful connection between iterative prompt lengthening and increasing task specificity. However, the formal contribution is the centerpiece: the paper claims to prove a specificity crossover theorem, and that theorem rests entirely on Proposition 3.8. As detailed below, Proposition 3.8 is not a valid information-theoretic lower bound, and the definition of task specificity is not well-posed as stated. The case studies and qualitative discussion may still support the position, but the advertised formal proof does not go through. The paper needs a substantially reworked formal framework, or a reframing of the theorem as a conjecture rather than a proven result.

major comments (4)
  1. [§3.5, Proposition 3.8] The lower bound L_NL(σ) ≥ σ + Δtrans is not an information-theoretic consequence. The proof claims that because I(Y;U) ≥ σ, the description U 'must carry at least σ bits about Y,' and encoding in natural language adds Δtrans. The first step can be made valid only in the sense that H(U) ≥ I(Y;U), so a source-coding argument gives L ≥ H(U) ≥ σ. The second step has no basis: Δtrans = D_KL(pM ∥ pNL) is a distributional divergence between the output distribution and a language-model distribution, not a mandatory per-message cost imposed on any description U. A concrete counterexample is σ=0: a constant or empty prompt achieves I(Y;Yhat)=0 with expected length 0, while the bound predicts L_NL(0) ≥ Δtrans > 0 whenever pM ≠ pNL. Thus the inequality is false as stated, and Theorem 3.10, which is derived by rearranging this inequality, is not established.
  2. [§3.3, Definition 3.4] The definition of task specificity σϵ(T) as 'the minimal mutual information I(Y;Yhat) such that Yhat∈[T] with probability > 1−ϵ' is not well-posed. Taken literally, for any task with more than one acceptable output one can choose Yhat to be a deterministic function of T supported on [T], making Yhat conditionally independent of the exact value of Y given T; the mutual information I(Y;Yhat) can then be driven to or near zero. The definition therefore collapses to σ=0 for essentially all non-singleton tasks, contradicting the intended 'specificity ladder.' To capture the intended notion—how much information about the user's envisioned Y must be transmitted to obtain an acceptable output—the definition needs an explicit fidelity or distortion constraint, e.g., a bound on E[d(Y,Yhat)] or a shrinking equivalence class around Y. Without such a constraint, the theorem's central variable σ is no
  3. [§3.6, Theorem 3.10 and Remark 3.11] Even if Proposition 3.8 were accepted, the theorem's content is largely a rearrangement of definitions: σ* is obtained by solving the inequality σ + Δtrans < H(Y) + R_formal. The only substantive non-definitional input is Remark 3.11's assumption that R_formal < Δtrans, which is asserted on the basis of the intuition that domain-specific languages are 'better optimized' for their domains. No measurement, empirical estimate, or derivation of Δtrans or R_formal is provided for any concrete domain. Consequently, the threshold σ* cannot be computed or tested, and the theorem is unfalsifiable in its current form. The authors should either supply an operational way to estimate these quantities or explicitly present the crossover as a qualitative conjecture supported by the case studies.
  4. [§3.2 and §3.5, formal-path modeling] The comparison is asymmetric by construction: the formal path is defined to always specify the exact output Y, with cost L_formal = H(Y) + R_formal independent of σ, while the natural-language path is allowed to exploit the equivalence class [T] and reduce information transmission to σ. This asymmetry is a modeling choice, not a theorem. In practice, formal languages can also underspecify—through parametrized functions, nondeterminism, abstraction, or incomplete specifications—and the cost of formal specification should then depend on σ as well. The paper does not justify why the formal path must operate only at full specification. As a result, the crossover conclusion is partly baked into the definitions, and the framework does not establish an intrinsic property of formal vs. natural language.
minor comments (6)
  1. [§4, Quantitative Illustration] The text reads: 'The total number of possible images is 256^{3×250×250} = 21,500,000.' If the intended expression is 2^{1,500,000}, the superscript formatting is missing and the equality is false as written; with the given alphabet argument, N ≥ 300,000 follows from 2^{5N} ≥ 2^{1,500,000}. Please correct the notation.
  2. [§3.2, Markov chain] The Markov chain Y→T→U→Yhat is stated without specifying the conditional distributions or the stochasticity of the generation model. Since I(Y;Yhat) is central to the definition of σ, these distributions should be made explicit.
  3. [§3.5, Definition 3.7] The 'expected description length' L_NL(σ) is not defined with respect to a distribution over the user's utterance strategy or a coding scheme (e.g., prefix-free codes). Source-coding lower bounds require such a specification; please clarify.
  4. [Appendix B, Lemma 3.6] The notation JSD(D) is used without a formal definition. The proof correctly derives the identity, but the Jensen-Shannon divergence for a set of distributions should be defined explicitly.
  5. [§3.6, Remark 3.12] If L_NL already denotes actual expected description length, the remark that 'natural language also has redundancy' is confusing: redundancy in natural-language descriptions would already be reflected in L_NL and should not be added on top of the lower bound. Clarify the intended role of this remark.
  6. [§4.1, Figure 3 caption] The caption mentions 'Gemini-3' without a citation or version detail. Please clarify whether this refers to a specific model release and add an appropriate reference or rephrase.

Circularity Check

2 steps flagged · score 7.0 of 10

The crossover threshold σ* is an algebraic rearrangement of an unproved lower bound (Prop 3.8) that already assumes the additive translation-gap cost; the central 'prediction' reduces to its own modeling assumption.

  1. other [§3.5, Proposition 3.8 and proof]
    "By (2), I(Y;U) ≥ I(Y;\hat{Y}) ≥ σ. The description U must therefore carry at least σ bits about Y. Encoding this in natural language incurs the translation gap, giving the bound."

    The additive lower bound L_NL(σ) ≥ σ + Δ_trans is asserted, not derived. Δ_trans = D_KL(p_M∥p_NL) is a divergence between output distributions, not a per-message length overhead; no source-coding argument connects I(Y;U) ≥ σ to σ + Δ_trans bits of description length. The bound even fails at σ=0, where a constant prompt achieves I(Y;Y_hat)=0 at zero length while Δ_trans > 0. Because the proof merely stipulates the term it later treats as a proved cost, the bound is an input in the form of a proposition.

  2. self definitional [§3.6, Theorem 3.10 proof]
    "By 3.8, achieving σ requires L_NL(σ) ≥ σ + Δ_trans. Formal specification costs L_formal = H(Y) + R_formal regardless of σ. Setting L_NL < L_formal yields: σ + Δ_trans < H(Y) + R_formal ⇒ σ < σ*"

    The threshold σ* = H(Y) + R_formal − Δ_trans is obtained by solving the inequality that Prop 3.8 already assumed. The theorem's upper regime ('natural language cannot be more efficient') is exactly the assumed lower bound rewritten as σ ≥ H(Y)+R_formal−Δ_trans. No independent measurement or derivation fixes Δ_trans or R_formal; Remark 3.11 supplies only 'intuitively, we expect'. The 'prediction' is therefore equivalent by construction to the additive bound chosen as input.

full rationale

The paper's central derivation chain is: define L_NL(σ) as minimum length for mutual information σ; assert (Prop 3.8) that this length is at least σ + Δ_trans; then solve the inequality against L_formal = H(Y)+R_formal to get σ* (Thm 3.10). The only non-definitional step is Prop 3.8, and its proof does not derive the Δ_trans term—it simply states that 'encoding this in natural language incurs the translation gap'. That is a modeling ansatz, not an information-theoretic consequence. Consequently the crossover theorem's content is contained in the assumed bound, and the paper's own Remark 3.11 concedes the threshold is meaningful only under an unmeasured inequality R_formal < Δ_trans. The paper does not fit parameters to data and call them predictions, nor does it rely on a uniqueness theorem; the circularity is of the self-definitional/algebraic kind. The self-citation to Don-Yehiya et al. (2023) (co-authored by Abend) is used only as supporting empirical evidence, not as load-bearing for the theorem, so it does not independently raise the score. Overall score 7: the central crossover result reduces by construction to an asserted lower bound, while the surrounding qualitative discussion and case studies are independent content.

Assumptions & free parameters 2 free parameters · 7 assumptions · 3 invented entities

The framework introduces three unmeasured quantities—task specificity, translation gap, and formal redundancy—and the crossover theorem depends on their relative size. The paper supplies no empirical estimate or measurement protocol for these quantities; the only substantive constraint is the expectation that R_formal < Δtrans.

free parameters (2)
  • Translation gap Δtrans
    Defined as D_KL(pM∥pNL) but never estimated for any domain; the argument requires Δtrans > 0 and Δtrans > R_formal.
  • Formal-language redundancy R_formal
    Defined as L_formal − H(Y) but never measured; Remark 3.11 assumes it is smaller than Δtrans.
assumptions (7)
  • domain assumption Natural language is optimized for underspecification in open-ended contexts (Zipf; Gibson et al.; Piantadosi et al.).
    Foundational premise from linguistics used throughout §2.
  • domain assumption Shorter utterances are more efficient (Goodman & Frank; Gibson et al.).
    Cost model in §3; justifies using description length as the efficiency metric.
  • domain assumption The user's intended output Y is a sample from a fixed prior pM over meaning space M, and tasks are constraint sets.
    Definitions 3.1–3.3 in §3.1–3.2; a modeling idealization that excludes contradictory or unrealizable intents.
  • domain assumption The formal path Y→Ŷ directly specifies the full output with expected length L_formal = H(Y) + R_formal, independent of task specificity.
    Definition 3.9 in §3.5; ignores partial formal specifications and learning costs, which are only partially acknowledged in Remark 3.14.
  • standard math Data Processing Inequality applies to the Markov chain Y → T → U → Ŷ.
    Equation (2) in §3.2; used to derive the natural-language cost lower bound.
  • standard math Uniform prior over a finite set of domain distributions in Lemma 3.6.
    Appendix B, equation (8); needed for the JSD decomposition.
  • ad hoc to paper Expectation that R_formal < Δtrans for a meaningful crossover.
    Remark 3.11 asserts this criterion is satisfied when R_formal is small and Δtrans > 0, but this ordering is not derived from data or prior literature.
invented entities (3)
  • Task specificity σ_ϵ(T)
    purpose: Central scalar measuring how much information about the intended output the task requires.
    Defined via minimal mutual information with probability 1−ε, but no measurement protocol or empirical estimate is given.
  • Translation gap Δtrans
    purpose: KL divergence from the domain distribution to the natural-language distribution, used as the overhead cost of natural-language specification.
    Introduced in Definition 3.5; no estimates are provided and the distribution pNL is left unspecified.
  • Formal-language redundancy R_formal
    purpose: Overhead of formal specification beyond the information-theoretic minimum H(Y).
    Defined in Definition 3.9 but never measured; its assumed ordering relative to Δtrans carries the crossover result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Position: Natural Language Should Not Fully Replace Formal Languages." pith.science (2026). https://pith.science/paper/QXGQGTQG

@misc{pith2026260720432,
  author       = {Pith},
  title        = {Pith review of: Position: Natural Language Should Not Fully Replace Formal Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QXGQGTQG}},
  note         = {Machine review of arXiv:2607.20432}
}
read the original abstract

Recent advances in large language models and their widespread adoption have prompted claims that natural language could entirely replace formal languages, such as programming languages for software design. In this position paper, we argue that this perspective overlooks fundamental linguistic properties of natural language, specifically that it is optimized for underspecification in open-ended contexts. We introduce a formal framework centered on *task specificity*, defining it as the information-theoretic reduction of uncertainty in an output space -- such as all possible images -- given a user's specific requirements. We prove a *specificity crossover theorem*, showing the existence of a threshold beyond which the cost to express formal requirements into natural language exceeds the cost of direct formal specification. By analyzing case studies across modalities, such as image generation, code synthesis, and audio production, we demonstrate that natural language excels at low specificity tasks, while formal languages are advantageous on tasks with stricter requirements. We conclude that natural and formal languages are complementary tools and advocate the development of hybrid systems that allow users to move across the specificity spectrum.

Figures

Figures reproduced from arXiv: 2607.20432 by the authors.

Figure 1
Figure 1. The Ladder of Specification: from genuine indifference (top) to overspecification (bottom), with corresponding compres￾sion benefits and error risks. Relaxing constraints over possible meaning is termed se￾mantic underspecification. In semantics, multiple frame￾works attempt to represent the “shared” meaning of a sen￾tence through a single, compact meta-structure (Reyle, 1993; Copestake et al., 2005). From a cogniti… view at source ↗
Figure 2
Figure 2. Two specification paths. Natural language path (solid blue): Y → T → U → Yˆ . The user forms a task T (allowing indifferent compression), describes it in natural language U (allow￾ing pragmatic compression), and the model generates Yˆ . Formal language path (dashed red): Y → Yˆ via direct specification. (i.e., expects in this context) but not specified in u. Not satisfying Cimp is considered a failure. A constraint … view at source ↗
Figure 3
Figure 3. Illustration of task specificity in image generation. As the description becomes longer, the space of images is narrower. At the limit, if a specific image is desired, perfectly describing it in words is inefficient (compared to filming or drawing). All images were generated by Gemini-3, except for the image from Disney’s 1996 101 Dalmatians (right). into full specification (2) or even overspecification (1), where u… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

104 extracted references · 4 canonical work pages

  1. [1]

    2018 , eprint=

    Studying the Difference Between Natural and Programming Language Corpora , author=. 2018 , eprint=

    2018

  2. [2]

    European Semantic Web Conference , pages=

    Bridging the gap between formal languages and natural languages with zippers , author=. European Semantic Web Conference , pages=. 2016 , organization=

    2016

  3. [3]

    , title =

    Dijkstra, Edsger W. , title =. Program Construction, International Summer Schoo , pages =. 1978 , isbn =

    1978

  4. [4]

    Bell system technical journal , volume=

    Prediction and entropy of printed English , author=. Bell system technical journal , volume=. 1951 , publisher=

    1951

  5. [5]

    The Twelfth International Conference on Learning Representations , year=

    Language Modeling Is Compression , author=. The Twelfth International Conference on Learning Representations , year=

  6. [6]

    1969 , publisher=

    Speech acts: An essay in the philosophy of language , author=. 1969 , publisher=

    1969

  7. [7]

    A Survey on LLM s for Story Generation

    Teleki, Maria and Bengali, Vedangi and Dong, Xiangjue and Janjur, Sai Tejas and Liu, Haoran and Liu, Tian and Wang, Cong and Liu, Ting and Zhang, Yin and Shipman, Frank and Caverlee, James. A Survey on LLM s for Story Generation. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.750

    doi:10.18653/v1/2025.findings-emnlp.750 2025

  8. [8]

    Controllable Text Summarization: Unraveling Challenges, Approaches, and Prospects - A Survey

    Urlana, Ashok and Mishra, Pruthwik and Roy, Tathagato and Mishra, Rahul. Controllable Text Summarization: Unraveling Challenges, Approaches, and Prospects - A Survey. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.93

    work page doi:10.18653/v1/2024.findings-acl.93 2024

Show all 104 references
  1. [9]

    2025 , eprint=

    R ^2 : A LLM Based Novel-to-Screenplay Generation Framework with Causal Plot Graphs , author=. 2025 , eprint=

    2025

  2. [10]

    Navigating the Path of Writing: Outline-guided Text Generation with Large Language Models

    Lee, Yukyung and Ka, Soonwon and Son, Bokyung and Kang, Pilsung and Kang, Jaewook. Navigating the Path of Writing: Outline-guided Text Generation with Large Language Models. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Comput...

    2025 doi

  3. [11]

    2026 , eprint=

    Lessons from the Field: An Adaptable Lifecycle Approach to Applied Dialogue Summarization , author=. 2026 , eprint=

    2026

  4. [12]

    Assessing the Performance of AI-Generated Code: A Case Study on GitHub Copilot , year=

    Li, Shuang and Cheng, Yuntao and Chen, Jinfu and Xuan, Jifeng and He, Sen and Shang, Weiyi , booktitle=. Assessing the Performance of AI-Generated Code: A Case Study on GitHub Copilot , year=

  5. [13]

    2025 , eprint=

    NaturalEdit: Code Modification through Direct Interaction with Adaptive Natural Language Representation , author=. 2025 , eprint=

    2025

  6. [14]

    CoRR , year=

    Symbolic Music Generation with Fine-grained Interactive Textural Guidance , author=. CoRR , year=

  7. [15]

    IEEE Access , year=

    Music generation using deep learning and generative AI: a systematic review , author=. IEEE Access , year=

  8. [16]

    Computer , volume=

    A specifier's introduction to formal methods , author=. Computer , volume=. 2002 , publisher=

    2002

  9. [17]

    arXiv preprint arXiv:2512.07702 , year=

    Guiding What Not to Generate: Automated Negative Prompting for Text-Image Alignment , author=. arXiv preprint arXiv:2512.07702 , year=

  10. [18]

    european conference on computer vision , pages=

    Understanding the Impact of Negative Prompts: When and How Do They Take Effect? , author=. european conference on computer vision , pages=. 2024 , organization=

    2024

  11. [19]

    and Kee, Kerk and Zampieri, Marcos and Calyam, Prasad , title =

    Oruche, Roland and Goruganthu, Sai Keerthana and Akula, Rithika and Cheng, Xiyao and Goni, Ashraful Md and Shibo, Bruce W. and Kee, Kerk and Zampieri, Marcos and Calyam, Prasad , title =. ACM Comput. Surv. , month = may, articleno =. 2025 , issue_date =. doi:10.1145/3729220 , ...

    2025 doi

  12. [20]

    Communications of the ACM , volume=

    Natural language dialogue for personalized interaction , author=. Communications of the ACM , volume=. 2000 , publisher=

    2000

  13. [21]

    , author=

    Speaker knowledge influences the comprehension of pragmatic inferences. , author=. Journal of Experimental Psychology: Learning, Memory, and Cognition , volume=. 2012 , publisher=

    2012

  14. [22]

    Nouns are Vectors, Adjectives are Matrices: Representing Adjective-Noun Constructions in Semantic Space

    Baroni, Marco and Zamparelli, Roberto. Nouns are Vectors, Adjectives are Matrices: Representing Adjective-Noun Constructions in Semantic Space. Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing. 2010

    2010

  15. [23]

    Cognitive science , volume=

    Composition in distributional models of semantics , author=. Cognitive science , volume=. 2010 , publisher=

    2010

  16. [24]

    Frontiers in Communication , volume=

    Information theory as a bridge between language function and language form , author=. Frontiers in Communication , volume=. 2022 , publisher=

    2022

  17. [25]

    Journal of semantics , volume=

    Dealing with ambiguities by underspecification: Construction, representation and deduction , author=. Journal of semantics , volume=. 1993 , publisher=

    1993

  18. [26]

    Research on language and computation , volume=

    Minimal recursion semantics: An introduction , author=. Research on language and computation , volume=. 2005 , publisher=

    2005

  19. [27]

    Language and linguistics compass , volume=

    Semantic underspecification in language processing , author=. Language and linguistics compass , volume=. 2009 , publisher=

    2009

  20. [28]

    arXiv preprint cmp-lg/9505034 , year=

    Semantic ambiguity and perceived ambiguity , author=. arXiv preprint cmp-lg/9505034 , year=

    arXiv

  21. [29]

    Logic, language and computation , volume=

    Dynamics and the semantics of dialogue , author=. Logic, language and computation , volume=

  22. [30]

    2000 , publisher=

    Presumptive meanings: The theory of generalized conversational implicature , author=. 2000 , publisher=

    2000

  23. [31]

    Semantics and Pragmatics , volume=

    Free choice permission as resource-sensitive reasoning , author=. Semantics and Pragmatics , volume=

  24. [32]

    Language , pages=

    Pragmatic halos , author=. Language , pages=. 1999 , publisher=

    1999

  25. [33]

    2007 , publisher=

    Approximate interpretation of number words , author=. 2007 , publisher=

    2007

  26. [34]

    2008 , publisher=

    Thoughts and utterances: The pragmatics of explicit communication , author=. 2008 , publisher=

    2008

  27. [35]

    1996 , publisher=

    Using language , author=. 1996 , publisher=

    1996

  28. [36]

    1986 , publisher=

    Relevance: Communication and cognition , author=. 1986 , publisher=

    1986

  29. [37]

    2007 , publisher=

    Bayesian rationality: The probabilistic approach to human reasoning , author=. 2007 , publisher=

    2007

  30. [38]

    The Journal of Logic Programming , volume=

    A survey of complexity results for non-monotonic logics , author=. The Journal of Logic Programming , volume=. 1993 , publisher=

    1993

  31. [39]

    1987 , school=

    A default logic approach to the derivation of natural language presuppositions , author=. 1987 , school=

    1987

  32. [40]

    Artificial intelligence , volume=

    A logic for default reasoning , author=. Artificial intelligence , volume=. 1980 , publisher=

    1980

  33. [41]

    2009 , publisher=

    The cultural origins of human cognition , author=. 2009 , publisher=

    2009

  34. [42]

    1998 , publisher=

    The symbolic species: The co-evolution of language and the brain , author=. 1998 , publisher=

    1998

  35. [43]

    , biburl =

    Zipf, George K. , biburl =

  36. [44]

    T he W orld of an O ctopus: H ow R eporting B ias I nfluences a L anguage M odel ' s P erception of C olor

    Paik, Cory and Aroca-Ouellette, St \'e phane and Roncone, Alessandro and Kann, Katharina. T he W orld of an O ctopus: H ow R eporting B ias I nfluences a L anguage M odel ' s P erception of C olor. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Pro...

    2021 doi

  37. [45]

    Coding theorems for a discrete source with a fidelity criterion , author=. IRE Nat. Conv. Rec , volume=

  38. [46]

    The Bell system technical journal , volume=

    A mathematical theory of communication , author=. The Bell system technical journal , volume=. 1948 , publisher=

    1948

  39. [47]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Sketchedit: Mask-free local image manipulation with partial sketches , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  40. [48]

    Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    AI Art and its Impact on Artists , author=. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    2023

  41. [49]

    Artificial intelligence review , volume=

    Artificial intelligence in the creative industries: a review , author=. Artificial intelligence review , volume=. 2022 , publisher=

    2022

  42. [50]

    Companion Publication of the 2024 Conference on Computer-Supported Cooperative Work and Social Computing , pages=

    Exploring the impact of AI-generated image tools on professional and non-professional users in the art and design fields , author=. Companion Publication of the 2024 Conference on Computer-Supported Cooperative Work and Social Computing , pages=

    2024

  43. [51]

    Patterns , volume=

    A decade of gender bias in machine translation , author=. Patterns , volume=. 2025 , publisher=

    2025

  44. [52]

    Transactions of the Association for Computational Linguistics , volume=

    Gender bias in machine translation , author=. Transactions of the Association for Computational Linguistics , volume=. 2021 , publisher=

    2021

  45. [53]

    Advances in Neural Information Processing Systems , volume=

    Optimizing prompts for text-to-image generation , author=. Advances in Neural Information Processing Systems , volume=

  46. [54]

    An invitation to cognitive science: Language , volume=

    Lexical semantics and compositionality , author=. An invitation to cognitive science: Language , volume=. 1995 , publisher=

    1995

  47. [55]

    Topoi , volume=

    The principle of semantic compositionality , author=. Topoi , volume=. 1994 , publisher=

    1994

  48. [56]

    2023 IEEE 31st International Conference on Network Protocols (ICNP) , pages=

    Rate-distortion-perception theory for semantic communication , author=. 2023 IEEE 31st International Conference on Network Protocols (ICNP) , pages=. 2023 , organization=

    2023

  49. [57]

    2021 IEEE International Symposium on Information Theory (ISIT) , pages=

    A rate-distortion framework for characterizing semantic information , author=. 2021 IEEE International Symposium on Information Theory (ISIT) , pages=. 2021 , organization=

    2021

  50. [58]

    Software: Practice and Experience , volume=

    Lightweight validation of natural language requirements , author=. Software: Practice and Experience , volume=. 2002 , publisher=

    2002

  51. [59]

    Program Verification: Fundamental Issues in Computer Science , pages=

    On formalism in specifications , author=. Program Verification: Fundamental Issues in Computer Science , pages=. 1993 , publisher=

    1993

  52. [60]

    Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Frontiers , pages=

    Drawing and Sketching: Art, Psychology, and Computer Graphics , author=. Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Frontiers , pages=

  53. [61]

    2021 , eprint=

    Evaluating Large Language Models Trained on Code , author=. 2021 , eprint=

    2021

  54. [62]

    Daniel Jurafsky and James H. Martin. Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition, with Language Models. 2025

    2025

  55. [63]

    Mind , volume=

    On concept and object , author=. Mind , volume=. 1951 , publisher=

    1951

  56. [64]

    Scientific American , volume=

    The origin of speech , author=. Scientific American , volume=. 1960 , publisher=

    1960

  57. [65]

    Cognitive science , volume=

    The cultural evolution of structured languages in an open-ended, continuous world , author=. Cognitive science , volume=. 2017 , publisher=

    2017

  58. [66]

    2025 , eprint=

    The Hidden Risks of LLM-Generated Web Application Code: A Security-Centric Evaluation of Code Generation Capabilities in Large Language Models , author=. 2025 , eprint=

    2025

  59. [67]

    ACM Computing Surveys (CSUR) , volume=

    A survey of machine learning for big code and naturalness , author=. ACM Computing Surveys (CSUR) , volume=. 2018 , publisher=

    2018

  60. [68]

    arXiv preprint arXiv:2301.11325 , year=

    Musiclm: Generating music from text , author=. arXiv preprint arXiv:2301.11325 , year=

    arXiv

  61. [69]

    Language to Logical Form with Neural Attention

    Dong, Li and Lapata, Mirella. Language to Logical Form with Neural Attention. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2016. doi:10.18653/v1/P16-1004

    2016 doi

  62. [70]

    Proceedings of the 2022 CHI conference on human factors in computing systems , pages=

    Design guidelines for prompt engineering text-to-image generative models , author=. Proceedings of the 2022 CHI conference on human factors in computing systems , pages=

    2022

  63. [71]

    Behaviour & Information Technology , volume=

    A taxonomy of prompt modifiers for text-to-image generation , author=. Behaviour & Information Technology , volume=. 2024 , publisher=

    2024

  64. [72]

    Human Learning by Model Feedback: The Dynamics of Iterative Prompting with Midjourney

    Don-Yehiya, Shachar and Choshen, Leshem and Abend, Omri. Human Learning by Model Feedback: The Dynamics of Iterative Prompting with Midjourney. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.253

    2023 doi

  65. [73]

    arXiv preprint arXiv:2204.06125 , volume=

    Hierarchical text-conditional image generation with clip latents , author=. arXiv preprint arXiv:2204.06125 , volume=

    arXiv

  66. [74]

    Advances in neural information processing systems , volume=

    Photorealistic text-to-image diffusion models with deep language understanding , author=. Advances in neural information processing systems , volume=

  67. [75]

    Information Technology & People , volume=

    AI in software programming: understanding emotional responses to GitHub Copilot , author=. Information Technology & People , volume=. 2025 , publisher=

    2025

  68. [76]

    Proceedings of the 26th International Conference on Intelligent User Interfaces , pages=

    Perfection not required? Human-AI partnerships in code translation , author=. Proceedings of the 26th International Conference on Intelligent User Interfaces , pages=

  69. [77]

    arXiv preprint arXiv:2310.06770 , year=

    Swe-bench: Can language models resolve real-world github issues? , author=. arXiv preprint arXiv:2310.06770 , year=

    arXiv

  70. [78]

    2012 , publisher=

    Mathematical methods in linguistics , author=. 2012 , publisher=

    2012

  71. [79]

    Cross-Task Generalization via Natural Language Crowdsourcing Instructions

    Mishra, Swaroop and Khashabi, Daniel and Baral, Chitta and Hajishirzi, Hannaneh. Cross-Task Generalization via Natural Language Crowdsourcing Instructions. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022. do...

    2022 doi

  72. [80]

    Trends in cognitive sciences , volume=

    How efficiency shapes human language , author=. Trends in cognitive sciences , volume=. 2019 , publisher=

    2019

  73. [81]

    1999 , publisher=

    Elements of information theory , author=. 1999 , publisher=

    1999

  74. [82]

    Proceedings of the ACM on Programming Languages , volume=

    Grounded copilot: How programmers interact with code-generating models , author=. Proceedings of the ACM on Programming Languages , volume=. 2023 , publisher=

    2023

  75. [83]

    Rissanen, J. J. , journal=. Generalized Kraft Inequality and Arithmetic Coding , year=

  76. [84]

    arXiv preprint arXiv:2406.00515 , year=

    A survey on large language models for code generation , author=. arXiv preprint arXiv:2406.00515 , year=

    arXiv

  77. [85]

    ACM Transactions on Software Engineering and Methodology , year=

    A survey on llm-based code generation for low-resource and domain-specific programming languages , author=. ACM Transactions on Software Engineering and Methodology , year=

  78. [86]

    ACM Computing Surveys , volume=

    A systematic survey of text summarization: From statistical methods to large language models , author=. ACM Computing Surveys , volume=. 2025 , publisher=

    2025

  79. [87]

    ACM Computing Surveys , volume=

    Survey on the evaluation of generative models in music , author=. ACM Computing Surveys , volume=. 2025 , publisher=

    2025

  80. [88]

    Neurocomputing , volume=

    Open-world story generation with structured knowledge enhancement: A comprehensive survey , author=. Neurocomputing , volume=. 2023 , publisher=

    2023

  81. [89]

    2025 , eprint=

    Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity , author=. 2025 , eprint=

    2025

  82. [90]

    Communications of the ACM , volume=

    On the naturalness of software , author=. Communications of the ACM , volume=. 2016 , publisher=

    2016

  83. [91]

    ACM Computing Surveys , volume=

    A survey on employing large language models for text-to-sql tasks , author=. ACM Computing Surveys , volume=. 2025 , publisher=

    2025

  84. [92]

    Trends in cognitive sciences , volume=

    Pragmatic language interpretation as probabilistic inference , author=. Trends in cognitive sciences , volume=. 2016 , publisher=

    2016

  85. [93]

    Science , volume=

    Predicting pragmatic reasoning in language games , author=. Science , volume=. 2012 , publisher=

    2012

  86. [94]

    Science , volume=

    Kinship categories across languages reflect general communicative principles , author=. Science , volume=. 2012 , publisher=

    2012

  87. [95]

    Cognitively Informed

    Zaslavsky, Noga and Kemp, Charles and Regier, Terry and Tishby, Naftali , Title =. Cognitively Informed

  88. [96]

    Journal of The Royal Society Interface , volume=

    The evolution of lossy compression , author=. Journal of The Royal Society Interface , volume=. 2017 , publisher=

    2017

  89. [97]

    Semantics and pragmatics , volume=

    Information structure: Towards an integrated formal theory of pragmatics , author=. Semantics and pragmatics , volume=

  90. [98]

    Synthese , volume=

    Vagueness, truth and logic , author=. Synthese , volume=. 1975 , publisher=

    1975

  91. [99]

    Syntax and semantics , volume=

    Logic and conversation , author=. Syntax and semantics , volume=

  92. [100]

    Annual Review of Linguistics , volume=

    The rational speech act framework , author=. Annual Review of Linguistics , volume=. 2023 , publisher=

    2023

  93. [101]

    Cognition , volume=

    The communicative function of ambiguity in language , author=. Cognition , volume=. 2012 , publisher=

    2012

  94. [102]

    arXiv preprint arXiv:2403.14072 , year=

    A taxonomy of ambiguity types for nlp , author=. arXiv preprint arXiv:2403.14072 , year=

    arXiv

  95. [103]

    Psychology of Learning and motivation , volume=

    Ambiguity, accessibility, and a division of labor for communicative success , author=. Psychology of Learning and motivation , volume=. 2008 , publisher=

    2008

  96. [104]

    Dealing with Semantic Underspecification in Multimodal NLP

    Pezzelle, Sandro. Dealing with Semantic Underspecification in Multimodal NLP. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023. doi:10.18653/v1/2023.acl-long.675

    2023 doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.