Austria

This paper introduces a pioneering methodology, termed StructTuning, to efficiently transform foundation Large Language Models (LLMs) into domain specialists. It significantly reduces the training corpus needs to a mere 5% while achieving an impressive 100% of traditional knowledge injection performance. Motivated by structured human education, we propose a novel two-stage strategy for knowledge injection and alignment: Structure-aware Continual Pre-Training (SCPT) and Structure-aware Supervised Fine-Tuning (SSFT). In the SCPT phase, we automatically extract the domain knowledge taxonomy and reorganize the training corpora, enabling LLMs to effectively link textual segments to targeted knowledge points within the taxonomy. In the SSFT phase, we explicitly prompt models to elucidate the underlying knowledge structure in their outputs, leveraging the structured domain insight to address practical problems. Our ultimate method was extensively evaluated across model architectures and scales on LongBench and MMedBench datasets, demonstrating superior performance against other knowledge injection methods. We also explored our method’s scalability across different training corpus sizes, laying the foundation to enhance domain-specific LLMs with better data utilization.

ACL 2025

Structure-aware Domain Knowledge Injection for Large Language Models

structured knowledge

knowledge injection

large language models

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

technical paper

Social determinants of health (SDOH) extraction from clinical text is critical for downstream healthcare analytics. Although large language models (LLMs) have shown promise, they may rely on superficial cues leading to spurious predictions. Using the MIMIC portion of the SHAC (Social History Annotation Corpus) dataset and focusing on drug status extraction as a case study, we demonstrate that mentions of alcohol or smoking can falsely induce models to predict current/past drug use where none is present, while also uncovering concerning gender disparities in model performance. We further evaluate mitigation strategies—such as prompt engineering and chain-of-thought reasoning—to reduce these false positives, providing insights into enhancing LLM reliability in health domains.

Spurious Correlations and Beyond: Understanding and Mitigating Shortcut Learning in SDOH Extraction with Large Language Models

Model merging for Large Language Models (LLMs) directly fuses the parameters of different models finetuned on various tasks, creating a unified model for multi-domain tasks. However, due to potential vulnerabilities in models available on open-source platforms, model merging is susceptible to backdoor attacks. In this paper, we propose $\textit{Merge Hijacking}$, the first backdoor attack targeting model merging in LLMs. The attacker constructs a malicious upload model and releases it. Once a victim user merges it with any other models, the resulting merged model inherits the backdoor while maintaining utility across tasks. Merge Hijacking defines two main objectives—effectiveness and utility—and achieves them through four steps. Extensive experiments demonstrate the effectiveness of our attack across different models, merging algorithms, and tasks. Additionally, we show that the attack remains effective even when merging real-world models. Moreover, our attack demonstrates robustness against two inference-time defenses (Paraphrasing and CLEANGEN) and one training-time defense (Fine-pruning).

Merge Hijacking: Backdoor Attacks to Model Merging of Large Language Models

The adeptness of Large Language Models (LLMs) in comprehending and following natural language instructions is critical for their deployment in sophisticated real-world applications. Existing evaluations mainly focus on fragmented constraints or narrow scenarios, but they overlook the comprehensiveness and authenticity of constraints from the user's perspective. To bridge this gap, we propose CFBench, a large-scale Chinese Comprehensive Constraints Following Benchmark for LLMs, featuring 1,000 curated samples that cover more than 200 real-life scenarios and over 50 NLP tasks. CFBench meticulously compiles constraints from real-world instructions and constructs an innovative systematic framework for constraint types, which includes 10 primary categories and over 25 subcategories, and ensures each constraint is seamlessly integrated within the instructions. To make certain that the evaluation of LLM outputs aligns with user perceptions, we propose an advanced methodology that integrates multi-dimensional assessment criteria with requirement prioritization, covering various perspectives of constraints, instructions, and requirement fulfillment. Evaluating current leading LLMs on CFBench reveals substantial room for improvement in constraints following, and we further investigate influencing factors and enhancement strategies. The data and code will be made available.

CFBench: A Comprehensive Constraints-Following Benchmark for LLMs

In recent research on large language models (LLMs), there has been a growing emphasis on aligning these models with human values to reduce the impact of harmful content. However, current alignment methods often rely solely on singular forms of feedback, such as preferences, annotated labels, or natural language critiques, overlooking the potential advantages of combining these feedback types. This limitation leads to suboptimal performance, even when ample training data is available. In this paper, we introduce Diverse AI Feedback (DAIF) as a novel method to enhance LLM alignment with diverse feedback mechanisms. Our approach involves collecting three distinct types of feedback tailored to problems of varying difficulty levels within the training dataset. Specifically, we exploit critique feedback for easy problems, refinement feedback for medium problems, and preference feedback for hard problems. By training our model with this diversified feedback, we achieve enhanced alignment performance while using less training data. To assess the effectiveness of DAIF, we evaluate it against previous methods in three downstream tasks: question answering, dialog generation, and text summarization. Experimental results demonstrate that DAIF achieves superior performance even with a smaller training dataset.

Diverse AI Feedback For Large Language Model Alignment

The sparsity of labelled data is an obstacle to the development of Relation Extraction (RE) models and the completion of databases in various biomedical areas. While being of high interest in drug-discovery, the literature on natural-products, reporting the identification of potential bioactive compounds from organisms, is a concrete example of such an overlooked topic. To mark the start of this new task, we created the first curated evaluation dataset and extracted literature items from the LOTUS database to build training sets. To this end, we developed a new sampler, inspired by diversity metrics in ecology, named Greedy Maximum Entropy sampler (https://github.com/idiap/gme-sampler). The strategic optimization of both balance and diversity of the selected items in the evaluation set is important given the resource-intensive nature of manual curation. After quantifying the noise in the training set, in the form of discrepancies between the text of input abstracts and the expected output labels, we explored different strategies accordingly. Framing the task as an end-to-end Relation Extraction, we evaluated the performance of standard fine-tuning (BioGPT, GPT-2 and Seq2rel) and few-shot learning with open Large Language Models (LLaMA 7B-65B). In addition to their evaluation in few-shot settings, we explore the potential of open LLMs as synthetic data generators and propose a new workflow for this purpose. All evaluated models exhibited substantial improvements when fine-tuned on synthetic abstracts rather than the original noisy data. We provide our best performing (f1-score=59.0) BioGPT-Large model for end-to-end RE of natural-products relationships along with all the training and evaluation datasets. See more details at https://github.com/idiap/abroad-re.

Relation Extraction in Underexplored Biomedical Domains: A Diversity-optimized Sampling and Synthetic Data Generation Approach

One of the major outstanding questions in computational semantics is how humans integrate the meaning of individual words into a sentence in a way that enables understanding of complex and novel combinations of words, a phenomenon known as compositionality. Two major approaches to modelling the process of compositionality are vector-based models, in which the meaning of a sentence is represented as a vector of numbers, and syntax-based models, in which the meaning of a sentence is represented as a structured tree of labelled components. A major barrier in assessing and comparing these contrasting approaches is the lack of large, relevant datasets for model comparison. This paper aims to address this gap by introducing a new dataset, STS3k, which consists of 2,800 pairs of sentences rated for semantic similarity by human participants. The sentence pairs have been selected to systematically vary different combinations of words, providing a rigorous test and enabling a clearer picture of the comparative strengths and weaknesses of vector-based and syntax-based methods. Our results show that when tested on the new STS3k dataset, state-of-the-art transformers poorly capture the pattern of human semantic similarity judgments, while even simple methods for combining syntax- and vector-based components into a novel hybrid model yield substantial improvements. We further show that this improvement is due to the ability of the hybrid model to replicate human sensitivity to specific changes in sentence structure. Our findings provide evidence for the value of integrating multiple methods to better reflect the way in which humans mentally represent compositional meaning.

Compositionality and Sentence Meaning: Comparing Semantic Parsing and Transformers on a Challenging Sentence Similarity Dataset

The brain's ability to perform complex computations at varying timescales is crucial, ranging from understanding single words to grasping the overarching narrative of a story. Recently, multi-timescale long short-term memory (MT-LSTM) models (Mahto et al. 2020; Jain et al. 2020) have been introduced to use temporally-tuned parameters to induce sensitivity to different timescales of language processing (i.e. related to near/distant words). However, whether the temporally-tuned information processing in MT-LSTMs is similar to the brain has not been explored using high temporal resolution recording modalities such as electroencephalography (EEG).To bridge this gap, we used an EEG dataset recorded while participants listened to Chapter 1 of "Alice in Wonderland" and trained ridge regression models to predict the temporally-tuned MT-LSTM embeddings from EEG responses. Our analysis reveals that EEG signal segments can effectively predict MT-LSTM embeddings across various timescales. For longer timescales, predictions are significant within an extended time window of Â±2 s around word onset, while for shorter timescales, significant predictions are confined to a narrow window ranging from -180 ms to 790 ms. Intriguingly, we observed that short timescale information is not only processed in the vicinity of word onset but also at distant time points. % , specifically at -1 and +2 s.These observations underscore the parallels and discrepancies between computational models and the neural mechanisms of the brain. As word embeddings are used more as \textit{in silico} models of semantic representation in the brain, the ability to induce word embeddings, tuned to different timescales, allows for an ever-increasing range of questions in language processing in humans and machines to be addressed via brain-model analyses.

Exploring Temporal Sensitivity in the Brain Using Multi-timescale Language Models: An EEG Decoding Study

This study explores the cognitive mechanisms underlying human language acquisition through grammar induction by a minimal cognitive architecture, with a short and flexible sequence memory as its most central feature. We use reinforcement learning for the task of identifying sentences in a stream of words from artificial languages. Results demonstrate the modelâs ability to identify frequent and informative multi-word chunks, reproducing characteristics of natural language acquisition. The model successfully navigates varying degrees of linguistic complexity, exposing efficient adaptation to combinatorial challenges through the reuse of sequential patterns. The emergence of parsimonious tree structures suggests an optimization for the sentence identification task, balancing economy and information. The cognitive architecture reflects aspects of human memory systems and decision-making processes, enhancing its cognitive plausibility. While the model exhibits limitations in generalization and semantic representation, its minimalist nature offers insights into some fundamental mechanisms of language learning. Our study demonstrates the power of this simple architecture and stresses the importance of sequence memory in language learning. Since other animals do not seem to have faithful sequence memory, this may be a key to understanding why only humans have developed complex languages.

Usage-based Grammar Induction from Minimal Cognitive Principles

Humans acquire their native languages by actively taking part in communicative interactions with their caregivers. These interactions are meaningful, intentional and situated in their everyday environment. The situated and communicative nature of the interactions is essential to the language acquisition process, as language learners depend on clues provided by the communicative environment to make sense of the utterances they perceive. As such, the linguistic knowledge they build up is rooted in linguistic forms, their meaning and their communicative function. When it comes to machines, the situated, communicative and interactional aspects of language learning are often passed over. This applies in particular to today's large language models (LLMs), where the input is purely text-based, and where the distribution of character groups or words serves as a basis for modelling the meaning of linguistic expressions. In this paper, we argue that this design choice lies at the root of a number of important limitations, in particular regarding the data hungriness of the models, their limited ability to perform logical and pragmatic reasoning, and their susceptibility to biases. At the same time, we make a case for an alternative approach that models how artificial agents can acquire linguistic structures by participating in situated communicative interactions. Through a selection of experiments, we show how the linguistic knowledge that is captured in the resulting models is of a fundamentally different nature than the knowledge captured by LLMs, and argue that this change of perspective provides a promising path towards more human-like language processing in machines.

Humans Learn Language from Situated Communicative Interactions. What about Machines?

Current open-domain neural semantics parsers show impressive performance. However, closer inspection of the symbolic meaning representations they produce reveals significant weaknesses: sometimes they tend to merely copy character sequences from the source text to form symbolic concepts, defaulting to the most frequent word sense based in the training distribution. By leveraging the hierarchical structure of a lexical ontology, we introduce a novel compositional symbolic representation for concepts based on their position in the taxonomical hierarchy. This representation provides richer semantic information and enhances interpretability. We introduce a neural "taxonomical" semantic parser Â to utilize this new representation system of predicates, and compare it with a standard neural semantic parser trained on the traditional meaning representation format, employing a novel challenge set and evaluation metric for evaluation. Our experimental findings demonstrate that the taxonomical model, trained on much richer and complex meaning representations, is slightly subordinate in performance to the traditional model using the standard metrics for evaluation, but outperforms it when dealing with out-of-vocabulary concepts. We further showed through neural model probing that training on a taxonomic representation enhances the model's ability to learn the taxonomical hierarchy. This finding is encouraging for research in computational semantics that aims to combine data-driven distributional meanings with knowledge-based symbolic representations.

Downloads

Next from ACL 2025

Spurious Correlations and Beyond: Understanding and Mitigating Shortcut Learning in SDOH Extraction with Large Language Models

Similar lecture

Memorizing is Not Enough: Deep Knowledge Injection Through Reasoning