Attempt to Differentiate: NLP, Computational Linguistics, & Digital Humanities

Intro: “ACL is not an AI conference”

“ACL is not an AI conference.” At ACL 2024, held in Bangkok, Thailand, this year’s ACL Chair, Emily M. Bender, put forth a highly controversial conclusion in her address.

She emphasized that ACL, as the Annual Meeting of the Association for Computational Linguistics, is not primarily focused on artificial intelligence, but rather on language technology and computational linguistics. While she acknowledged that machine learning (including deep learning) has provided many useful tools for these two, she argued that problems arise when the focus shifts towards AI.

Bender believes that ACL’s focus should be on “language,” grounded in an understanding of data, including knowledge of how language works (i.e., linguistics) and dataset documentation. On the other hand, AI focuses on “how to endow machines with human-like intelligence.”

Given the booming development of AI, its commercialization, and negative trends in the research community—such as the improper use of benchmarks, requirements to compare with closed SOTA models, and the excessive size of datasets leading to a lack of retained data — Bender remarked that if your research question revolves around “how to prove that my machine is intelligent,” such a focus could distort research practices. The upsurging emphasis on AI has also led to poor review practices, where papers that do not employ LLMs or fail to deliver LLM results at the SOTA scale may be deemed uninteresting.

Bender asserted that ACL should be a space focused on language technology, fostering interdisciplinary research, a field that cares about linguistic communities, and a place where we can rationally discuss the societal impacts of our research and technology.

This view sparked significant controversy. Some argued that the two fields have organically merged and cannot be separated; others felt this stance reflected narrow-mindedness; still, some expressed understanding, agreeing that AI does not necessarily emphasize language itself.

Something Personal

Where exactly are the boundaries of disciplines?

How Enigmas Developed

Being an art student through my entire high school, receiving only humanities training for four years consecutively, and attending a STEM university to study linguistics, I have truly endured the low status of my major — this is the first one—no further comment about that.

After ChatGPT went viral, the gap between the humanities and sciences has unprecedentedly widened (I mean, sure, there has for long been a gap, but not one so enormous that left an art student no choice but to go STEM 😇), and humanities jobs have been replaced by a colossal margin — this is the second.

Amid the devastating impact of this second point, a new term, “digital humanities,” emerged from a small number of successful initiatives led by non-humanities scholars. Out of nowhere, China’s most prestigious universities have launched relevant journals, organized conferences, and even established “social science laboratories.” But what exactly is digital humanities? As a linguistics student at a science university, this question has indeed perplexed me for a long time.

Is it the deciphering of oracle bone inscriptions that can be published in Nature? Or the digitization and intelligent annotation of ancient texts? Or is it simply discourse analysis supported by corpus data and fancy visualizations? We saw suddenly, a wave of computer scientists entered, but the ensuing work was all taken up by traditional humanities scholars, struggling to survive under the tsunami of the information age. What exactly is digital humanities? No one knows. This is the third.

Also after overcoming the devastating blow of the second point, I started learning mathematics, statistics, programming, and machine learning from scratch. Gradually, I became involved in research related to LLMs, and I realized how interesting and meaningful it is to apply LLMs to so many tasks, some of which definitely have something to do with linguistics — this is the fourth.

The Question: NLP vs. CL

When I first entered this field and was preparing to choose my graduate direction, I became confused: What is the difference between Natural Language Processing (NLP) and Computational Linguistics (hereinafter referred to as CL)?

In terms of content, or even the names of top NLP conferences (ACL, duh), these two fields seem to be identical. However, whether in previous booms of deep learning models or in the current AI wave, it seems to be purely a celebration of “NLP,” while linguistics and computational linguistics have faded into obscurity.

Yet, after a whole fruitful year of reading more papers and gaining exposure to new advancements in both NLP technology and linguistic theory, I’ve reached a conclusion:

The previously dominant approach — represented by deep learning, big data, weak rules, empiricism, and connectionism — will no longer hold its supremacy. With the paradigm shift brought by LLMs, methods based on small data, strong rules, rationalism, and symbolic systems will experience a revival.

Inspired by: 杨延宁.(2024).语言学与人工智能：下一个十年.外语教学理论与实践(03),1-9.

The disappearance of the term “computational linguistics” was a manifestation of empiricism’s dominance, but with the resurgence of rationalism, the boundaries between NLP and CL will blur, leading to a gradual elevation of the status of both linguistics and CL.

Agents’ Reincarnation

Here, I’d like to give credit to some representative works around Agents (智能体) which have effortlessly given rise to symbolism, rationalism, or what I’d like to put it, the linguistic future of AI.

Agent is nothing novel. Thanks to an online presentation delivered by Guohao Li (李国豪) from CAMEL AI, I got to realize what was the opposite of neural networks, which rescued the entire AI community from the condemned, stagnant hot water of SVM and Random Forest, and how it illuminates against the current surge of AI kicked off by LLMs. I even purchased The Society of Mind by Marvin Minsky he mentioned, discovering a radically different school (学派) of AI - Symbolic AI, which featured symbols, natural language logic, and human-like reasonings, from the 1980s.

Guohao Li and Bernard Ghanem. CAMEL: Communicative Agents for “Mind” Exploration of Large Language Model Society.

Consequently, multi-agent collaboration comes into play.

Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1–22, San Francisco CA USA. ACM.

Based on this triumphant “believable social behavior simulation” around LLMs, another paradigm focusing on task finishing emerged.

Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. 2024. ChatDev: Communicative Agents for Software Development. arXiv:2307.07924 [cs].

Junkai Li, Siyu Wang, Meng Zhang, Weitao Li, Yunghwei Lai, Xinhui Kang, Weizhi Ma, and Yang Liu. 2024. Agent Hospital: A Simulacrum of Hospital with Evolvable Medical Agents. arXiv:2405.02957 [cs].

Some even more entrepreneurial attempts barged into my vision as well - using 1,000 LLM-based agents in Minecraft to start an autonomous civilization, with religion, politics, and governance made by themselves from scratch! That’s when people know, that Agents are coming back.

So is expertise in linguistics. I grinned.

What If We Differentiate Them Anyway?

From a practical perspective, and as you may have noticed from my previous explanation, NLP and CL have already become an integrated whole (I intentionally avoided using these two terms just now). However, in terms of terminology, there should be a distinction between them. This is a distinction that requires a “deliberate effort to differentiate” in order to be discerned. Most of the time, such a distinction is unnecessary, but to resolve my confusion, it is essential.

So let’s do this.

AI and LLMs pierce through all three fields, with:

• NLP focusing more on technical aspects like deep learning and algorithms, often with less contribution from linguistics.
• CL, on the other hand, integrates linguistic methods to address computational challenges related to language, bridging technical and theoretical approaches.
• DH applies computational techniques to support traditional humanities, aiming to enhance research in these areas.
• There surely exist hazy areas and works around them. Therefore, the spectrum from technical to humanities is highlighted along the bottom axis.

This is original. I personally don’t feel 100% qualified to exert this taxonomy to the general NLP/CL/DH community, so might as well take it as, again, something personal.

…and here is a spreadsheet focusing more on other details…

Among them, “The First Principle Thinking” was inspired by someone I knew from an international program to Cambridge. Lovely place btw. With nothing theoretical about “Symbolic AI” on his mind, this is what he came up with the minute I described how an “irony” is composed in the light of linguistics, given the context that he did an irony detection system in said program.

The following might be irrelevant, but I’d like to distinguish these bad boys as well. They caused both trouble and strength during my exploration getting rid of that existential crisis. (aforementioned point two)

Outro: But Does It Matter?

It does NOT.

As it turned out, ChatGPT has really opened a chapter for a former art student to delve into CL, LLMs, and NLP (no DH ‘cuz it’s just too close to home, lol 😂), with a more effortless, dummy-friendly, intuitive, elegant yet powerful way to learn everything it requires along the way - Python, PyTorch, CUDA, Linux, Fine-tuning, and so on. What matters most is the fact that I have long moved past the paranoia of thinking “linguistics is just useless, going CS is all I need”, and getting ready to pursue something I love, identify with, and feel a “sense of acquisition” about. Whatever the name was, is, or will be.

Maybe just not labeling anything. How about more “I do it”?