Natural Language Understanding: Decoding Thought

Natural language understanding stands at the very intersection of artificial intelligence, theoretical linguistics, and cognitive science, addressing the formidable challenge of enabling computational systems to comprehend human discourse. While machines have long excelled at manipulating syntactic structures and computing statistical distributions of tokens, the genuine apprehension of underlying semantic nuance, contextual pragmatics, and intentionality represents a foundational milestone in computational intelligence.

Automated Natural Language Understanding

1. Concise Definition

Automated Natural Language Understanding (NLU) is a specialized subdiscipline of natural language processing (NLP) and artificial intelligence dedicated to the computational comprehension of spoken, written, or multimodal human language. Rather than merely processing text via statistical tokenization or surface-level syntactic patterns, NLU systems translate unstructured linguistic inputs into structured, machine-interpretable semantic representations capable of supporting reasoning, inference, and task execution.

At its core, automated NLU seeks to bridge the profound chasm between linguistic expression and conceptual meaning. By modeling phonological, morphological, lexical, syntactic, semantic, and discourse-level features, an NLU pipeline enables autonomous systems to discern not only what words are uttered, but what propositional content, functional intent, and affective coloration those words communicate within a situated communicative context.

In technical architectures, automated NLU operates as an inferential engine that converts ambiguous, elliptical, and context-dependent human text into explicit symbolic logic, frame representations, or high-dimensional latent vectors. Consequently, it forms the critical cognitive backbone of conversational interfaces, knowledge-graph extractors, automated diagnostic agents, and automated reasoning platforms.

2. Etymology and Linguistic Origin

The phrase natural language understanding synthesizes three distinct Latinate and Germanic linguistic lineages that converge upon modern cognitive computing. The adjective natural derives via Old French from the Latin naturalis (originating from natura, meaning “birth, character, or the essential properties of things”), used in technical computer science to differentiate organic human languages (such as English, Arabic, Mandarin, or Quechua) from engineered formal systems (such as C++, Python, or propositional calculi).

The noun language originates from the Old French langage, stemming directly from the Latin lingua, signifying “tongue, speech, or dialect.” Finally, the verbal noun understanding traces back to the Old English understandan (“to comprehend, perceive, or grasp the meaning of”), constructed from the prefix under- (meaning “between” or “among,” cognate with the Latin inter-) and standan (“to stand”). Historically, to understand meant to stand firmly among concepts, apprehending their mutual relationships.

The term formally coalesced within artificial intelligence during the late 1960s and early 1970s. Early pioneers, seeking to distinguish systems that merely manipulated strings from those that genuinely comprehended internal state models, adopted the term to designate systems capable of deep natural language understanding. Landmark efforts, such as Terry Winograd’s SHRDLU program (1971), explicitly adopted the label to signal that machines were modeling micro-worlds rather than calculating shallow surface statistics.

3. Pronunciation and Grammatical Form

Pronunciation: /ˌnætʃ.ər.əl ˈlæŋ.ɡwɪdʒ ˌʌn.dərˈstæn.dɪŋ/
Part of Speech: Compound noun phrase (uncountable).
Common Abbreviations and Variations: NLU; Automated NLU; Computational Natural Language Comprehension.
Usage Context: Typically functions as a singular non-count substantive subject or direct object in scientific discourse (e.g., “Automated natural language understanding enables the system to infer conversational implicature”). It can also function adjectivally in compound attributions, such as “NLU benchmarks,” “NLU architectures,” or “NLU pipelines.”

4. Detailed Conceptual Explanation

The scope of automated natural language understanding encompasses the automated conversion of unbounded human linguistic variation into deterministic or probabilistic representations of meaning. Human language is inherently underspecified, highly ambiguous, context-dependent, and reliant upon vast reservoirs of shared unstated common-sense knowledge. Consequently, NLU must solve problems that extend far beyond syntactic parsing or pattern recognition.

To achieve comprehension, an NLU architecture must manage linguistic processing across several hierarchically stratified strata:

First, lexical and morphological disambiguation requires discerning word sense in context. A word such as “bank” acquires completely divergent denotations depending on whether the surrounding utterance involves a river, a financial transaction, a billiard shot, or an aeronautical maneuver. NLU algorithms utilize contextual embeddings, frame repositories, or graph ontology mappings to resolve these polysemous variations.

Second, syntactic-semantic binding requires mapping grammatical arguments (such as subject, direct object, and prepositional phrases) onto semantic roles (such as Agent, Patient, Experiencer, Instrument, and Beneficiary). This process, known broadly as semantic parsing or semantic role labeling (SRL), allows a computer to determine not merely the grammatical structure of a sentence, but who performed what action to whom, when, where, why, and how.

Third, pragmatic and discourse-level analysis resolves reference, presupposition, and conversational implicature. Natural dialogue frequently employs anaphora and coreference (“Jane entered the room; she took her coat off”), requiring the system to identify that “Jane” and “she” represent identical real-world entities. Furthermore, pragmatics requires recognizing speech acts: recognizing that the query “Can you open the door?” is universally intended as an imperative request for physical action rather than an inquiry into the listener’s physical capabilities.

Finally, boundaries must be drawn between surface natural language processing and deep natural language understanding. While standard NLP encompasses tasks like optical character recognition, speech-to-text transcription, part-of-speech tagging, and statistical machine translation, NLU specifically demands an internal model of semantic content. An NLU system must possess the capacity to draw non-trivial inferences, answer counterfactual questions based on text, detect logical contradictions, and ground linguistic symbols into operational real-world actions or computational schemas.

5. Historical Development

The historical evolution of automated NLU reflects the broader developmental trajectory of artificial intelligence, traversing three major epochs: rule-based symbolic systems, empirical statistical approaches, and contemporary connectionist foundation models.

The symbolic paradigm dominated the field from the 1950s through the 1980s. Following Alan Turing’s foundational proposal of the imitation game, known as the Turing test, early researchers attempted to formalize comprehension through explicit rules. In the mid-1960s, Joseph Weizenbaum created ELIZA at MIT, which simulated understanding using pattern matching and keyword substitution, though it possessed zero actual comprehension. In contrast, Terry Winograd’s 1971 SHRDLU represented a monumental leap forward: it permitted a user to converse in English with a robotic arm manipulating geometric blocks inside a simulated “blocks world.” SHRDLU successfully understood complex syntactic clauses, remembered discourse history, and resolved ambiguities by checking physical feasibility. However, these early symbolic systems suffered from the “brittleness problem”; they operated exclusively in narrow microworlds and collapsed when encountering novel syntactic patterns or unexpected vocabulary.

The late 1980s and 1990s witnessed the statistical turn. Propelled by the availability of large electronic text corpora, such as the Penn Treebank, and advances in statistical machine learning, researchers abandoned hand-crafted formal grammars in favor of probabilistic models. Hidden Markov Models (HMMs), Maximum Entropy Markov Models (MEMMs), and Conditional Random Fields (CRFs) became the standard tools for information extraction, named-entity recognition, and shallow semantic parsing. Concurrently, lexical databases such as WordNet, developed by George Miller and Christiane Fellbaum at Princeton, provided formal ontologies of synsets, hypernyms, and hyponyms that empowered algorithmic disambiguation.

The contemporary era began with the deep learning revolution of the 2010s, progressing from distributed word representations (Word2Vec, GloVe) to recurrent neural networks (LSTMs) and, ultimately, to the Transformer architecture introduced by Vaswani et al. in 2017. Pre-trained bidirectional representations, epitomized by BERT (Devlin et al., 2018), and modern autoregressive large language models (such as GPT-4, Llama, and Claude) unified syntactic analysis, semantic parsing, and pragmatic reasoning into self-attention mechanisms trained on trillions of tokens. These systems demonstrate unprecedented cross-domain understanding, capable of passing professional examinations, generating creative analogies, and interpreting subtle conversational nuances.

6. Theoretical Foundations

Automated NLU does not inhabit an intellectual vacuum; it is anchored in formal linguistic theories, cognitive psychology, and mathematical logic.

One foundational paradigm is Montague Grammar and Formal Semantics, originated by logician Richard Montague in the 1970s. Montague famously posited that there is no important theoretical difference between the formal languages of logicians and the natural languages of human beings. Utilizing lambda calculus, intensional logic, and model theory, formal semantics models linguistic meaning as truth conditions: to understand the meaning of a declarative sentence is to know precisely the conditions under which that sentence is true.

A second major theoretical pillar is Frame Semantics, pioneered by linguist Charles Fillmore. Fillmore argued that words cannot be understood in isolation; instead, lexical tokens evoke structured conceptual schemas or “frames.” For instance, understanding the verb “to buy” inevitably evokes the “Commercial Transaction” frame, which inherently contains semantic slots for a Buyer, a Seller, Goods, and Money. This theoretical framework directly gave rise to FrameNet, a major computational resource extensively employed in training semantic parsing and slot-filling models.

A third theoretical root resides in the Distributional Hypothesis, attributed to linguists Zellig Harris and John Rupert Firth (“You shall know a word by the company it keeps”). This hypothesis claims that words with similar semantic distributions occur in similar linguistic environments. This principle underlies all modern vector-space models and neural embedding spaces: by projecting words, sentences, and documents into high-dimensional geometric spaces, semantic similarity corresponds directly to geometric proximity (e.g., cosine similarity), enabling computers to quantify conceptual relationships mathematically.

Finally, Cognitive Linguistics and Embodied Cognition, championed by George Lakoff and Mark Johnson, argue that human semantic comprehension is fundamentally grounded in sensorimotor experience and conceptual metaphors. This school of thought poses a persistent challenge to purely statistical or text-only NLU systems, arguing that genuine understanding cannot emerge purely from ungrounded token relationships without anchoring symbols in perceptual reality or physical embodiment.

7. Key Components, Types, and Dimensions

Modern automated natural language understanding architectures break down into specialized components, processing tasks, and organizational dimensions:

• Intent Recognition (Classification): The process of identifying the primary communicative goal or functional purpose behind an input utterance (e.g., classifying “Can you show me my account balance?” as CheckBalance).
• Named Entity Recognition (NER) and Slot Filling: The extraction of specific parametric entities within the discourse (e.g., tagging dates, geographical locations, personal names, monetary amounts, or technical codes) to populate operational computational variables.
• Semantic Role Labeling (SRL): Determining the predicate-argument structure of clauses, explicitly mapping who did what to whom, by what means, and in what context.
• Coreference Resolution: Determining when two or more linguistic expressions refer to the identical discourse referent (e.g., linking “the President,” “she,” and “Dr. Aris Thorne” within a multi-paragraph text).
• Sentiment and Stance Detection: Analyzing the affective valence, emotional orientation, subjectivity, and explicit ideological stance expressed by the speaker or writer toward a target subject.
• Textual Entailment and Natural Language Inference (NLI): Determining directional inferential relationships between a premise and a hypothesis, identifying whether the hypothesis is True (Entailment), False (Contradiction), or Undetermined (Neutral).
• Semantic Parsing: Converting informal, unstructured natural language queries into executable formal representations, such as SQL queries, SPARQL queries, or lambda-calculus expressions.
• Discourse Parsing and Pragmatics: Analyzing coherence relations between multiple sentences (e.g., Elaboration, Explanation, Contrast) and discerning implicit meaning via conversational context.

8. Examples and Illustrative Cases

To illuminate how automated natural language understanding operates in practical scenarios, consider several concrete illustrations across distinct functional domains.

Case 1: Conversational Banking and Slot Extraction
A user enters the following query into an automated virtual assistant: “Transfer four hundred dollars from my savings to my primary checking account tomorrow morning.”
An automated NLU engine decomposes this sentence through synchronized steps: Intent Classification tags the objective as Initiate_Funds_Transfer; Slot Filling identifies Source_Account: Savings, Target_Account: Checking_Primary, Amount: $400.00, and Execution_Time: Current_Date + 1_Day (08:00 AM). The engine resolves the temporal deictic expression “tomorrow morning” into an absolute timestamp by cross-referencing system clock metadata, producing an executable API payload without requiring human intervention.

Case 2: Resolving Syntactic and Semantic Ambiguity (The Winograd Schema)
Consider the paired sentences designed by Terry Winograd to test deep contextual reasoning:
Sentence A: “The trophy would not fit in the brown suitcase because it was too large.”
Sentence B: “The trophy would not fit in the brown suitcase because it was too small.”
A superficial syntactic parser fails here because the surface syntax is identical. However, an advanced NLU model resolves the pronoun “it” correctly in both cases: in Sentence A, common-sense spatial physics dictates that “it” refers to the trophy (the contained object being too large for the container); in Sentence B, the same physics dictates that “it” refers to the suitcase (the container being too small to accommodate the contained object).

Case 3: Automated Clinical Phenotyping
A physician writes an unstructured electronic health record (EHR) note: “Patient presented with acute dyspnea and bilateral lower extremity edema. Denies chest pain, fever, or prior pulmonary embolism. Started on intravenous furosemide.”
A clinical NLU system processes the narrative, mapping “acute dyspnea” to ICD-10 code R06.00 and “bilateral lower extremity edema” to R60.0. Crucially, the system uses negation detection algorithms (such as NegEx) to identify that “chest pain” and “pulmonary embolism” are explicitly negated, preventing false-positive diagnostic coding. It simultaneously links the prescription of furosemide to the probable clinical phenotype of congestive heart failure.

9. Measurement and Assessment

Evaluating automated natural language understanding requires rigorous psychometric and computational benchmarks designed to test specific facets of reasoning, robust generalization, and linguistic inference.

Historically, individual NLU components were evaluated using classification metrics: Precision (positive predictive value), Recall (sensitivity), and the harmonic mean of the two, the F1-score. For semantic parsing and slot-filling, exact-match accuracy of the generated abstract syntax tree or SQL statement serves as the standard criterion.

In 2018, the NLP research community transitioned to unified, holistic benchmarks. The GLUE (General Language Understanding Evaluation) benchmark, introduced by Wang et al., collected nine diverse sentence-level tasks, including sentiment analysis (SST-2), sentence similarity (STS-B), and natural language inference (MNLI, QNLI, RTE). As deep learning models quickly surpassed human performance baselines on GLUE, the more rigorous SuperGLUE benchmark was deployed, incorporating demanding tasks like multi-hop reading comprehension (MultiRC), causal reasoning (COPA), and complex coreference resolution (WSC).

Beyond discriminative benchmarks, modern evaluation employs comprehensive generative suites such as MMLU (Massive Multitask Language Understanding), which evaluates broad factual and academic understanding across 57 domains, and BIG-bench (Beyond the Imitation Game Benchmark), which includes tasks assessing irony detection, mathematical deduction, and theory of mind. Furthermore, adversarial evaluation suites, such as CheckList (Ribeiro et al., 2020), subject NLU systems to behavioral testing—including invariance tests (checking whether changing a name alters the prediction) and directional expectation tests (checking whether adding negative qualifiers correctly flips sentiment)—to detect hidden model brittle points.

10. Applications and Practical Significance

The societal and commercial deployment of automated NLU has transformed modern industry, governance, and daily human experience.

In enterprise operations, NLU powers intelligent document processing (IDP). Organizations automatically extract contract terms, liability clauses, and termination dates from millions of pages of legal documentation. In customer support, omni-channel conversational agents resolve inquiries autonomously by discerning user intent and emotional agitation, routing escalated cases to human agents alongside summarized contextual briefs.

In healthcare and biomedicine, NLU platforms parse scientific literature and electronic health records to support clinical trial matching, identify adverse drug events, and accelerate medical discovery. Systems indexing PubMed automatically construct biological knowledge graphs linking genes, diseases, and chemical compounds, identifying therapeutic pathways that would require decades of manual reading.

In public policy and defense, automated NLU facilitates open-source intelligence gathering, monitoring media streams across languages to detect emerging geopolitical crises, tracking disinformation campaigns, and evaluating societal sentiment during economic shocks or public health emergencies. Furthermore, accessibility technologies utilize NLU to generate dynamic descriptions of visual interfaces for visually impaired individuals, synthesize complex educational materials into simplified reading levels, and provide real-time translation that preserves pragmatic and cultural subtexts.

11. Research and Empirical Evidence

Decades of empirical experimentation have validated the cognitive efficacy and operational bounds of automated NLU architectures.

The foundational paper by Devlin et al. (2018) introducing BERT (Bidirectional Encoder Representations from Transformers) demonstrated empirically that bidirectional contextual pre-training generates fundamentally superior representations of linguistic meaning compared to unidirectional models. BERT established state-of-the-art results across eleven NLU benchmarks simultaneously, proving that context from both the left and right sides of a token is essential for deep semantic comprehension.

Subsequent empirical work by Brown et al. (2020) on GPT-3 documented the phenomenon of in-context few-shot learning. Their research revealed that scaling autoregressive language models by orders of magnitude (to 175 billion parameters) enabled systems to perform diverse NLU tasks—such as translation, question answering, and unscrambling words—purely through natural language prompting, without any downstream task-specific fine-tuning or parameter updates.

Concurrently, critical empirical work has highlighted persistent vulnerabilities. Ribeiro et al. (2020) demonstrated that state-of-the-art NLU models that achieved superhuman performance on GLUE frequently failed elementary behavioral tests; for example, changing a single neutral word in a sentence frequently caused sentiment classification systems to reverse their predictions. Similarly, research by McCoy et al. (2019) demonstrated that NLU models often rely on superficial statistical heuristics (such as lexical overlap) rather than genuine compositional logic to solve natural language inference tasks, exposing an ongoing divergence between statistical correlation and authentic human-level comprehension.

12. Cultural and Cross-Cultural Considerations

Automated natural language understanding faces extraordinary challenges when deployed across divergent linguistic systems and cross-cultural communicative paradigms.

First, the vast majority of historical NLU development has focused disproportionately on high-resource languages, particularly English, Mandarin, and major European tongues. Languages characterized as low-resource, which include thousands of African, indigenous American, and regional Asian languages, lack the digitized text corpora, annotated treebanks, and computational funding required to train high-parameter foundation models. This creates a severe global “digital divide” in automated comprehension technologies.

Second, cross-cultural pragmatics varies radically across human societies. Edward T. Hall’s sociological distinction between low-context and high-context cultures illustrates this friction. In low-context cultures (such as Germany or the United States), communication is direct, explicit, and literal; the semantic weight rests primarily within the uttered text itself. In high-context cultures (such as Japan or Arab societies), meaning is deeply embedded within interpersonal relationships, non-verbal hierarchy, shared history, and unstated social situational cues. An NLU model trained primarily on low-context text struggles profoundly to interpret high-context implicature, often misinterpreting deferential politeness markers or nuanced evasions as neutral or uninformative assertions.

Third, linguistic typologies introduce severe architectural challenges. Highly agglutinative languages (e.g., Turkish, Finnish, Hungarian) and polysynthetic languages construct complex composite concepts within a single morphological word, confounding standard tokenization schemes designed for isolating languages like English. Without specialized morphological decompounding, standard subword tokenizers create fragmented representations that obscure underlying semantic relationships.

13. Criticisms, Debates, and Limitations

Despite its remarkable technical advancements, automated NLU remains one of the most intellectually contested frontiers in modern cognitive science and artificial intelligence.

A primary philosophical objection is John Searle’s famous Chinese Room Argument (1980). Searle argued that a computer running formal syntactic programs to manipulate symbols does not possess genuine semantic understanding or intentionality. The machine simply manipulates ungrounded tokens based on structural rules; it does not understand Chinese in the manner that a conscious human understands language. In modern terms, critics argue that deep learning models are merely “stochastic parrots” (Bender et al., 2021)—sophisticated statistical mirrors that recombine memorized human linguistic forms without possessing authentic internal world models, beliefs, desires, or intentions.

A related theoretical debate concerns the Symbol Grounding Problem, articulated by Stevan Harnad (1990). Can an artificial neural network trained solely on text ever truly understand the meaning of physical words like “warmth,” “gravity,” “pain,” or “red”? Embodied cognition theorists argue that without sensorimotor interaction with an external physical environment, linguistic tokens remain ungrounded references circulating within a closed symbolic loop, incapable of reaching genuine semantic apprehension.

From an engineering standpoint, automated NLU systems suffer from severe practical limitations:

• Hallucination and Factuality: Generative NLU models regularly generate plausible-sounding, syntactically flawless assertions that are completely factually fictitious.
• Brittleness and Adversarial Vulnerability: Minor perturbations, typographic errors, or carefully engineered adversarial prompts can cause an otherwise robust NLU system to completely misinterpret user intent.
• Opacity and Lack of Interpretability: Deep neural models function largely as black boxes; determining the precise causal pathway through which an attention matrix arrived at a specific semantic interpretation remains an active, unresolved research problem.
• Common-Sense Reasoning Gaps: Current systems consistently struggle with basic physics, spatial intuition, and naive psychology that any five-year-old child navigates effortlessly.

14. Related Terms and Distinctions

To ensure academic precision, automated natural language understanding must be distinguished from several adjacent computational disciplines:

• Natural Language Processing (NLP): The overarching discipline encompassing all computational treatments of human language. NLP includes low-level syntactic tasks (tokenization, sentence boundary detection, lemmatization), speech processing, and document formatting, whereas NLU is strictly the semantic and pragmatic subset focused on extracting conceptual meaning.
• Natural Language Generation (NLG): The inverse of NLU. While NLU maps unstructured human language into structured computational representations (Text-to-Meaning), NLG translates structured computational data or semantic frames into coherent, natural human language (Meaning-to-Text).
• Optical Character Recognition (OCR): The mechanical or electronic conversion of images of typed, handwritten, or printed text into machine-encoded text. OCR operates purely at the visual perceptual layer and possesses no intrinsic comprehension of semantic content.
• Information Retrieval (IR): The science of searching for documents, metadata, or structured records within large unstructured collections (e.g., web search engines). Traditional IR relies upon keyword matching and inverted indices (such as TF-IDF or BM25), whereas modern semantic search integrates NLU to comprehend the underlying conceptual intent of a query.
• Automatic Speech Recognition (ASR): The acoustic transcription of spoken audio signals into written text strings. ASR provides the raw orthographic input to downstream NLU engines but does not inherently analyze or comprehend the linguistic meaning of the transcribed speech.

15. Summary and Key Takeaways

Automated Natural Language Understanding represents one of the crowning intellectual ambitions of computational intelligence: the transformation of ambiguous, culturally situated, and complex human discourse into structured, actionable semantic insight. Historically advancing from rigid, rule-based symbolic micro-worlds to flexible, high-parameter transformer foundation models, NLU unifies formal linguistic theory, distributional semantics, and deep neural architectures.

Modern NLU underpins critical applications across contemporary society, including enterprise information extraction, conversational virtual assistants, clinical informatics, and automated cross-lingual translation. However, substantial barriers persist. The digital divide between high-resource and low-resource languages, the challenge of high-context cross-cultural pragmatics, adversarial brittleness, and the profound theoretical hurdles of symbol grounding and semantic hallucination demonstrate that current computational systems remain imperfect emulators of human cognition.

In conclusion, while contemporary automated systems have achieved unprecedented operational mastery over syntactic form and broad empirical correlation, the ultimate goal of developing an autonomous artificial agent that fully comprehends the human condition, complete with common sense, perceptual embodiment, and authentic communicative intentionality, remains one of the preeminent scientific quests of the twenty-first century.

References

• Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM FAccT Conference (pp. 610–623). Association for Computing Machinery. https://doi.org/10.1145/3442188.3445922
• Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., … Amodei, D. (2020). Language models are few-shot learners. In Advances in Neural Information Processing Systems (Vol. 33, pp. 1877–1901). Curran Associates, Inc. https://proceedings.neurips.cc/paper/2020/hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html
• Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. https://arxiv.org/abs/1810.04805
• Fillmore, C. J. (1976). Frame semantics and the nature of language. Annals of the New York Academy of Sciences, 280(1), 20–32. https://doi.org/10.1111/j.1749-6632.1976.tb25467.x
• Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1–3), 335–346. https://doi.org/10.1016/0167-2789(90)90087-6
• McCoy, R. T., Pavlick, E., & Linzen, T. (2019). Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (pp. 3428–3448). Association for Computational Linguistics. https://doi.org/10.18653/v1/P19-1334
• Montague, R. (1970). Universal grammar. Theoria, 36(3), 373–398. https://doi.org/10.1111/j.1755-2567.1970.tb00434.x
• Ribeiro, M. T., Wu, T., Guestrin, C., & Singh, S. (2020). Beyond accuracy: Behavioral testing of NLP models with CheckList. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 4902–4912). Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-main.442
• Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417–424. https://doi.org/10.1017/S0140525X00005756
• Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing Systems (Vol. 30, pp. 5998–6008). Curran Associates, Inc. https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html
• Wang, A., Singh, A., Michael, J., Hill, F., Levy, O., & Bowman, S. R. (2018). GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP (pp. 353–355). Association for Computational Linguistics. https://doi.org/10.18653/v1/W18-5446
• Winograd, T. (1972). Understanding Natural Language. Academic Press. https://doi.org/10.1016/C2013-0-10499-5