The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Not by text prediction alone, according to Emily M. Bender and Alexander Koller. In their 2020 ACL position paper, they argue that a system trained only to predict strings can become exceptionally good at modeling linguistic form without gaining access to the relationship between an utterance, a speaker’s communicative intent, and the world being discussed. Fluent, useful output is therefore not, by itself, proof of human-like natural language understanding (NLU).
The argument is deliberately narrower than the claim that every language model is incapable of understanding. Bender and Koller present text-trained models as potentially useful components of a larger NLU system, while insisting that claims about “meaning” match the training signal and the evidence supplied by evaluations.
What is Bender and Koller’s central argument?
The paper’s thesis is that form and meaning should not be treated as interchangeable. A language model can learn regularities in the signals people produce—words, syntax, and patterns of co-occurrence—without necessarily learning why a person produced a particular signal or what it refers to outside the signal itself.
Recommended Free Tools
Bender and Koller focus on systems trained only through string prediction. Such a system receives linguistic material and is optimized to predict another piece of linguistic material. The objective can reward remarkably detailed knowledge of how language is used, but it does not directly provide the relation between a form and the communicative intent that gives that form meaning.
Popular on Amazon
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
The paper is a position piece, not a measurement study that assigns a score to “understanding.” Its warning is about how researchers describe capability: success on prediction or manipulation of language should not automatically be reported as evidence of human-like NLU.
What do “linguistic form” and “meaning” mean here?
Linguistic form is the observable signal: an utterance, written sentence, sequence of tokens, or other structure that can be recorded and processed. Linguistic meaning, in the authors’ working definition, is the relation between that form and communicative intent.
“We take (linguistic) meaning to be the relation between a linguistic form and communicative intent.”
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.
This definition puts meaning in a relationship rather than inside a string as a self-contained object. Consider the sentence “The meeting is tomorrow.” The same sequence of words can be a reminder, a correction, a threat to a deadline, or an answer to a scheduling question. The relevant interpretation depends on what the speaker is trying to accomplish and on the situation in which the sentence is used.
A model can learn that “meeting” often appears near “tomorrow,” and it can generate a plausible continuation. That statistical competence is evidence about its representation of form. It is not automatically evidence that the model has identified a speaker, an actual meeting, a calendar, or an intention to remind someone.
Also popular on Amazon
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
Why does string prediction create a form–meaning gap?
String prediction supplies an indirect signal. The model is penalized when its predicted token differs from the observed token, not when it misunderstands a speaker’s goal or misidentifies an object in the world. Large collections of text contain traces of how people use language, so a model can learn correlations that reflect aspects of meaning. Bender and Koller’s point is that those traces do not, through the text-only training signal alone, establish the underlying relation to the world and to communicative intent.
What the objective directly rewards
- Predicting likely words, tokens, or sequences in context.
- Capturing grammatical and stylistic regularities.
- Reproducing patterns associated with topics, genres, and discourse.
- Producing text that resembles the linguistic data used for training.
What the objective does not directly provide
- A connection between a referring expression and a particular object or event outside the text.
- Knowledge of why a speaker chose an utterance in a specific situation.
- Independent confirmation that a generated statement is true in the world.
- A guarantee that a successful prediction reflects the same kind of understanding a human listener uses.
The distinction is about the source of evidence, not about whether text can contain useful information. Text can encode descriptions, explanations, and accounts of experience. The authors argue that a form-only learning signal leaves the crucial relation underdetermined: many different situations or intentions can be compatible with similar linguistic patterns.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteDoes fluent output prove that a model understands language?
No. Fluency demonstrates that a system can produce well-formed or contextually plausible language; it does not settle whether the system has connected that language to communicative intent and the world. The paper asks readers to keep those claims separate when interpreting model behavior.
| Observation | What it is evidence for | What it does not establish on its own |
|---|---|---|
| High next-token or sequence-prediction performance | Skill at modeling regularities in linguistic form | Human-like understanding of reference, intention, or truth |
| Grammatical, coherent generated text | Ability to manipulate and produce familiar forms | That the system knows why an utterance is appropriate |
| Success on a language benchmark | Competence on the benchmark’s task and data distribution | A general proof of NLU across situations and communicative goals |
| Useful performance in an application | Practical value for that application | That the internal process is analogous to human understanding |
This does not make benchmarks or applications irrelevant. It changes what their results can legitimately support. A benchmark can show that a model performs a task under specified conditions; the stronger claim that the model understands language requires evidence tied to the relevant aspects of language use and grounding.
What kind of NLU do the authors have in mind?
Bender and Koller describe human-analogous NLU as involving more than the structure of language. It includes the use of language and the ability to ground language in the world. A full-scale system could therefore contain a language model while also obtaining information through interactions, perception, world knowledge, or other mechanisms that connect utterances to situations and intentions.
Also popular on Amazon
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
They are not proposing that every text-trained model has no value. Their conclusion allows that large models may learn “some reflection of meaning into the linguistic form,” which can be useful in applications. The caution is that reflected information should not be described as complete, human-like understanding without evidence for the missing connections.
Grounding is a requirement in the argument, not a product specification
The paper does not provide a shopping list of sensors, databases, or architectural modules that automatically produce NLU. “Grounding” names the needed connection between language and the world; the paper’s claim is that this connection is not supplied by string prediction alone. Designing and testing a grounded system remains a separate engineering and scientific problem.
How do the paper’s thought experiments clarify the issue?
The thought experiments make the form–meaning distinction easier to see by separating what is observable in language from what gives that language significance. At a high level, they ask readers to consider situations in which linguistic signals could be generated or interpreted without the system having access to the relevant world or communicative purpose.
If the same or similar strings are compatible with different underlying situations, then matching the strings cannot by itself identify which situation or intention is present. The example is not meant to deny that a model can infer useful regularities. It illustrates why those regularities should not be treated as conclusive evidence that the model has acquired the relation defined as meaning.
These are conceptual tests of an explanatory claim, not reported experiments with a named accuracy score. The paper uses them to challenge loose descriptions of model behavior, not to assert that language models fail every practical task.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Also popular on Amazon
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
What should an evaluator measure before claiming “understanding”?
The paper’s reporting lesson can be turned into a practical review checklist. Before using “understands,” identify what the system was trained to do, what the evaluation observes, and which part of language use the result supports.
- State the capability precisely. Say whether the result concerns prediction, classification, generation, dialogue behavior, or another defined task.
- Inspect the training signal. Record whether the system learned from strings alone or also received information that links language to situations, actions, or communicative goals.
- Define the intended meaning relation. Specify whether the claim concerns reference, speaker intent, truth, instructions, or another aspect of use.
- Test beyond surface form. Use cases in which similar wording has different implications, or different wording expresses the same goal, so that copying familiar patterns is insufficient.
- Check grounding and use. Evaluate whether the system’s interpretation remains appropriate when the relevant world state, participants, or task constraints change.
- Separate utility from explanation. Report that a system is useful when that is what the evidence shows; do not convert usefulness into a claim about human-like internal understanding.
- State the boundary of the result. Explain which situations, languages, domains, and communicative contexts were actually tested.
This framework does not demand one universal test. It demands that the evidence match the word being used. A result about linguistic form can be important even when it does not establish the broader meaning relation.
What exactly does “meaning cannot be be learned from form alone” claim?
In the conclusion, Bender and Koller write:
“In this paper, we have argued that in contrast to some current hype, meaning cannot be learned from form alone.”
Read in context, “form alone” refers to systems whose learning signal is only the linguistic signal and its prediction. The sentence is not a claim that no model can ever acquire useful semantic information, nor that a language model cannot participate in an NLU system. It is a claim that text-form prediction, by itself, does not supply sufficient evidence for the relation between language and communicative intent that the authors call meaning.
Free tools Windows power users keep installed
One-click scans. No signup required.
What did the 2020 paper establish, and what did it leave open?
| The paper supports | The paper does not settle |
|---|---|
| A distinction between linguistic form and meaning as a relation to communicative intent | A universal test that determines whether any contemporary model understands |
| A warning against treating string-prediction success as proof of NLU | Whether a particular later architecture acquires grounding through additional signals |
| The view that human-analogous NLU includes language use and grounding in the world | Which engineering design would be sufficient for full NLU |
| The possibility that large language models are useful components of larger systems | A numerical estimate of how much “meaning” any model has learned |
The work appeared in July 2020, before many subsequent model releases and evaluation debates. It should therefore be read as a conceptual argument about how to frame claims, not as a current survey of every model or as a final verdict on all research that followed.
Also popular on Amazon
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Why does this distinction matter in an age of abundant data?
More data and larger models can improve predictions of linguistic form, sometimes dramatically. That improvement can make generated language appear increasingly intentional, knowledgeable, or context-sensitive. The paper’s warning becomes more important as appearance and capability become harder to separate: readers need to know whether a system is demonstrating a measured task skill, exploiting regularities in its data, or showing evidence of a connection to a world and a speaker’s goal.
Precise language also improves engineering decisions. If a system is being used to draft text, retrieve information, or classify content, an evaluation of those functions may be enough. If it is expected to follow an instruction in a changing physical environment, infer an unfamiliar person’s intention, or make claims about the world, the evaluation must examine those relationships directly.
How should you read the paper’s contribution?
“Climbing towards NLU” is best understood as a correction to overconfident terminology. Bender and Koller do not dismiss language modeling; they ask researchers to describe what the training task and evaluation actually demonstrate. Their form–meaning distinction gives readers a disciplined way to interpret fluent output without assuming that linguistic resemblance equals understanding.
The paper’s enduring question is simple: what evidence connects a system’s manipulation of language to the intentions and world that language is used to communicate? Until an evaluation supplies that connection, the responsible conclusion is competence on the tested linguistic task—not proof of human-like NLU.
Frequently Asked Questions
Is Bender and Koller’s paper saying that language models are useless?
No. The authors explicitly allow that large language models can provide useful behavior and can serve as components in a broader NLU system. Their objection is to describing usefulness or fluency as settled evidence of human-like understanding.
Would adding non-text inputs automatically give a model meaning?
The paper says that grounding in the world is part of human-analogous NLU, but it does not claim that any particular combination of sensors, images, or databases is sufficient. Whether an added signal supplies the needed relation has to be demonstrated by the system’s tasks and evaluations.
Why is communicative intent part of the definition of meaning?
The same wording can perform different actions in different conversations, while different wording can serve the same goal. Focusing on intent prevents an evaluation from treating a sentence as meaningful merely because its tokens are statistically familiar.
Is this paper a benchmark or a survey of current AI models?
No. It is a conceptual position paper from the Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, published in July 2020. It argues for careful claims about NLU rather than reporting a cross-model leaderboard or a statistic that measures understanding.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
