When the LLM is abstract, what is the irreplaceable value of linguistics?

This article is part of the PRO Members Communication, and follows "PRO Members of Machine Heart" at the end of the article to read more on the topic.

In recent years, the development of large-scale language models has pushed natural language processing into a new phase. Through large-scale pre-training, models have been able to complete multiple language tasks such as translation, Q&A, writing, and code generation, and the research pathways used to build models for single tasks are being gradually replaced by general language models. As the model's capabilities continue to improve, researchers are no longer focusing on whether the model is able to complete a certain language task, but on further analyzing the mechanisms for forming the model's internal language capabilities, and the relationship between these capabilities and the use and understanding of human language.

Christopher Manning, a professor of computer science at Stanford University, discussed in a recent interview the changing direction of NLP research in the age of big models, arguing that the development of big models did not end language research, but pushed research away from the language processing capabilities themselves, and further into how language capabilities are formed and whether models have a deeper understanding capacity.

Directory

01. Why is it so hard for the LLM to grasp "the meaning of words"?
What has shifted the focus of NLP research away from mission capability?...

Why is it so hard for an LLM to grasp what it means to be foreign?

1 Christopher Manning, a professor of computer science at Stanford University, was recently interviewed about the ongoing impact of linguistic thought on the study of large language models, the three dimensions of language communication, language learning and linguistic meaning, the progress made in the study of large language models, and the current issues still being explored in the field of natural language processing (NLP).

1 Christopher Manning, professor of computer science at Stanford University and head of the Stanford NLP Group, has a long history of natural language processing, language representation learning, and deep learning research. He has been involved in writing the foundations of statistical natural language processing and promoting the development of NLP tools such as the Stanford CoreNLP.

2Christopher Manning notes that the development of large language models is changing the structure of problems in the field of natural language processing (NLP). Past research has focused on whether the model can accomplish specific language tasks, while current research has begun to focus further on whether the model really understands the role of language in communication.

1 Prior to the emergence of LLM, NLP studies often revolved around specific tasks such as machine translation, sentence analysis, text classification and questionnaires, with different tasks often requiring different models to be designed. With the development of Transformers and large-scale pre-training models, a unified model has been able to cover a large number of language tasks, prompting researchers to re-examine the mechanisms of language ability formation itself.

3· Manning points out that LLM's current advantages are mainly focused on language processing, whereas the language in real communication also contains information such as situations, goals, and the intention of the speaker. If the model cannot combine this information, it may not be possible to accurately judge what the user really wants to express based on the language's superficial laws alone.

1 In Manning's example of "it's a bit cold here", the sentence literally describes only the temperature, but in a specific communication environment, it can also mean that one wants the other to close the window or adjust the air. The text content of the sentence does not change, but the purpose of the expression changes as the situation changes.

4 As the big language model gradually addresses traditional NLP problems such as grammar, syntax, and basic semantic analysis, Manning believes that "Pragmatics", "Discourse", and "Dialogue" will become important directions for the big model to further understand language ability.

1 The current LLM has been able to generate consistent responses based on the above context, but real dialogue requires the model to continuously update the judgments of the user's objectives. For example, the user is constantly adding conditions in continuous communication, and the model needs to understand what information changes the task's objectives, rather than simply spelling out the previous content.

Manning describes the current LLM as a "language machine" that already has certain capabilities, but finds it difficult to properly express uncertainty and often gives definite answers when facts are insufficient, thereby affecting both sides of the conversation's judgments on the reliability of information.

1 The LLM has been able to identify many implicit intentions based on the above context, but Manning believes that there is still a problem of "overconfidence". Humans often use vague limiting words such as "maybe" and "I'm not so sure" to convey confidence when there is insufficient evidence, and models often give vague statements when there is insufficient evidence, interfering with both sides of the conversation's judgement of the reliability of information.

2 Manning notes that linguistics has in the past been called the "junk of linguistics and NLP" because many phenomena that could not be interpreted solely by sentences and semantic rules were eventually incorporated into linguistic research.

6·Manning's interpretation of discourse is the study of how language comprehension extends from a single sentence to a sequence of texts or phrases, establishing connections between different expressions and forming a whole meaning. In contrast to the analysis of the literal meaning of a sentence, discourse comprehension requires models to track people, events, references, and subject changes in context to determine how subsequent information relates to the previous content.

1 Manning uses the usual "this method" as an introduction to the paper. The paper often begins with "for when a new architecture is proposed" and concludes the next sentence with "this method improves previous performance problems" to understand the need to return the substitution back to a specific technique from the previous one, and then to put the conventional "performance problems" back into the corresponding experimental bottle.

2 For the LLM, the context-processing capability allows it to make use of large amounts of foreground information. However, in long text or continuous communication, the model still needs to judge which information has been established, which content has been modified by subsequent information, and what logical connections exist between different sentences.

7 Dialogue focuses on the interactive process of language exchange, i.e. how the two sides of the exchange continually confirm information through multiple expressions, coordinate intentions and advance common goals. Manning argues that dialogue is central to the study of multi-cycle coordination of intentions, i.e. whether the model can flexibly modify expression when there is a lack of understanding or information, and respond appropriately when there is a lack of information, unclear intentions or mutual misunderstanding.

1 Human daily communication has complex rules of interaction, which often actively reduce the degree of certainty of expression when the speaker lacks a grasp of something; and re-establish common understanding by supplementing explanation, confirmation or questioning when there is a bias in the understanding of the two sides. Manning believes that this more human-characteristic linguistic mechanism will continue to provide space for cross-sectional research in linguistics and machine learning.

From a linguistic point of view, why is a big language model an abstract learner?

Around the abstraction and learning ability of semantics in LLM, Manning shares their team's recent paper "Abstraction drives language learning", which argues that models form class abstraction very early in training, a result that tends to support the theoretical position of "abstract priority" in language learning.