Mihai Surdeanu

Oscar Sainz, Iker García-Ferrero, Alon Jacovi et al. · ibm-research

The 1st Workshop on Data Contamination (CONDA 2024) focuses on all relevant aspects of data contamination in natural language processing, where data contamination is understood as situations where evaluation data is included in pre-training corpora used to train large scale models, compromising evaluation results. The workshop fostered a shared task to collect evidence on data contamination in current available datasets and models. The goal of the shared task and associated database is to assist the community in understanding the extent of the problem and to assist researchers in avoiding reporting evaluation results on known contaminated resources. The shared task provides a structured, centralized public database for the collection of contamination evidence, open to contributions from the community via GitHub pool requests. This first compilation paper is based on 566 reported entries over 91 contaminated sources from a total of 23 contributors. The details of the individual contamination events are available in the platform. The platform continues to be online, open to contributions from the community.

Shahriar Golchin, Mihai Surdeanu, Nazgol Tavabi et al.

Domain adaptation for large neural language models (NLMs) is coupled with massive amounts of unstructured data in the pretraining phase. In this study, however, we show that pretrained NLMs learn in-domain information more effectively and faster from a compact subset of the data that focuses on the key information in the domain. We construct these compact subsets from the unstructured data using a combination of abstractive summaries and extractive keywords. In particular, we rely on BART to generate abstractive summaries, and KeyBERT to extract keywords from these summaries (or the original unstructured text directly). We evaluate our approach using six different settings: three datasets combined with two distinct NLMs. Our results reveal that the task-specific classifiers trained on top of NLMs pretrained using our method outperform methods based on traditional pretraining, i.e., random masking on the entire data, as well as methods without pretraining. Further, we show that our strategy reduces pretraining time by up to five times compared to vanilla pretraining. The code for all of our experiments is publicly available at https://github.com/shahriargolchin/compact-pretraining.

Shahriar Golchin, Mihai Surdeanu

Data contamination, i.e., the presence of test data from downstream tasks in the training data of large language models (LLMs), is a potential major issue in measuring LLMs' real effectiveness on other tasks. We propose a straightforward yet effective method for identifying data contamination within LLMs. At its core, our approach starts by identifying potential contamination at the instance level; using this information, our approach then assesses wider contamination at the partition level. To estimate contamination of individual instances, we employ "guided instruction:" a prompt consisting of the dataset name, partition type, and the random-length initial segment of a reference instance, asking the LLM to complete it. An instance is flagged as contaminated if the LLM's output either exactly or nearly matches the latter segment of the reference. To understand if an entire partition is contaminated, we propose two ideas. The first idea marks a dataset partition as contaminated if the average overlap score with the reference instances (as measured by ROUGE-L or BLEURT) is statistically significantly better with the completions from guided instruction compared to a "general instruction" that does not include the dataset and partition name. The second idea marks a dataset partition as contaminated if a classifier based on GPT-4 with few-shot in-context learning prompt marks multiple generated completions as exact/near-exact matches of the corresponding reference instances. Our best method achieves an accuracy between 92% and 100% in detecting if an LLM is contaminated with seven datasets, containing train and test/validation partitions, when contrasted with manual evaluation by human experts. Further, our findings indicate that GPT-4 is contaminated with AG News, WNLI, and XSum datasets.

Shahriar Golchin, Mihai Surdeanu

We propose the Data Contamination Quiz (DCQ), a simple and effective approach to detect data contamination in large language models (LLMs) and estimate the amount of it. Specifically, we frame data contamination detection as a series of multiple-choice questions, devising a quiz format wherein three perturbed versions of each instance, subsampled from a specific dataset partition, are created. These changes only include word-level perturbations. The generated perturbations, along with the original dataset instance, form the options in the DCQ, with an extra option accommodating the selection of none of the provided options. Given that the only distinguishing signal among the options is the exact wording with respect to the original dataset instance, an LLM, when tasked with identifying the original dataset instance, gravitates towards selecting the original one if it has been exposed to it. While accounting for positional biases in LLMs, the quiz performance reveals the contamination level for the tested model with the dataset partition to which the quiz pertains. Applied to various datasets and LLMs, under controlled and uncontrolled contamination, our findings, while fully lacking access to training data and model parameters, suggest that DCQ achieves state-of-the-art results and uncovers greater contamination levels through memorization compared to existing methods. Also, it proficiently bypasses more safety filters, especially those set to avoid generating copyrighted content.

Mohaddeseh Bastan, Mihai Surdeanu, Niranjan Balasubramanian

Natural language inference (NLI) is critical for complex decision-making in biomedical domain. One key question, for example, is whether a given biomedical mechanism is supported by experimental evidence. This can be seen as an NLI problem but there are no directly usable datasets to address this. The main challenge is that manually creating informative negative examples for this task is difficult and expensive. We introduce a novel semi-supervised procedure that bootstraps an NLI dataset from existing biomedical dataset that pairs mechanisms with experimental evidence in abstracts. We generate a range of negative examples using nine strategies that manipulate the structure of the underlying mechanisms both with rules, e.g., flip the roles of the entities in the interaction, and, more importantly, as perturbations via logical constraints in a neuro-logical decoding system. We use this procedure to create a novel dataset for NLI in the biomedical domain, called BioNLI and benchmark two state-of-the-art biomedical classifiers. The best result we obtain is around mid 70s in F1, suggesting the difficulty of the task. Critically, the performance on the different classes of negative examples varies widely, from 97% F1 on the simple role change negative examples, to barely better than chance on the negative examples generated using neuro-logic decoding.

Zheng Tang, Mihai Surdeanu

We propose an explainable approach for relation extraction that mitigates the tension between generalization and explainability by jointly training for the two goals. Our approach uses a multi-task learning architecture, which jointly trains a classifier for relation extraction, and a sequence model that labels words in the context of the relation that explain the decisions of the relation classifier. We also convert the model outputs to rules to bring global explanations to this approach. This sequence model is trained using a hybrid strategy: supervised, when supervision from pre-existing patterns is available, and semi-supervised otherwise. In the latter situation, we treat the sequence model's labels as latent variables, and learn the best assignment that maximizes the performance of the relation classifier. We evaluate the proposed approach on the two datasets and show that the sequence model provides labels that serve as accurate explanations for the relation classifier's decisions, and, importantly, that the joint training generally improves the performance of the relation classifier. We also evaluate the performance of the generated rules and show that the new rules are great add-on to the manual rules and bring the rule-based system much closer to the neural models.

Shahriar Golchin, Mihai Surdeanu, Nazgol Tavabi et al.

We propose a novel task-agnostic in-domain pre-training method that sits between generic pre-training and fine-tuning. Our approach selectively masks in-domain keywords, i.e., words that provide a compact representation of the target domain. We identify such keywords using KeyBERT (Grootendorst, 2020). We evaluate our approach using six different settings: three datasets combined with two distinct pre-trained language models (PLMs). Our results reveal that the fine-tuned PLMs adapted using our in-domain pre-training strategy outperform PLMs that used in-domain pre-training with random masking as well as those that followed the common pre-train-then-fine-tune paradigm. Further, the overhead of identifying in-domain keywords is reasonable, e.g., 7-15% of the pre-training time (for two epochs) for BERT Large (Devlin et al., 2019).

Zhengzhong Liang, Tushar Khot, Steven Bethard et al.

Considerable progress has been made recently in open-domain question answering (QA) problems, which require Information Retrieval (IR) and Reading Comprehension (RC). A popular approach to improve the system's performance is to improve the quality of the retrieved context from the IR stage. In this work we show that for StrategyQA, a challenging open-domain QA dataset that requires multi-hop reasoning, this common approach is surprisingly ineffective -- improving the quality of the retrieved context hardly improves the system's performance. We further analyze the system's behavior to identify potential reasons.

Shahriar Golchin, Mihai Surdeanu, Steven Bethard et al.

In-context learning (ICL) has proven to be an effective strategy for improving the performance of large language models (LLMs) with no additional training. However, the exact mechanism behind this performance improvement remains unclear. This study is the first to show how ICL surfaces memorized training data and to explore the correlation between this memorization and performance on downstream tasks across various ICL regimes: zero-shot, few-shot, and many-shot. Our most notable findings include: (1) ICL significantly surfaces memorization compared to zero-shot learning in most cases; (2) demonstrations, without their labels, are the most effective element in surfacing memorization; (3) ICL improves performance when the surfaced memorization in few-shot regimes reaches a high level (about 40%); and (4) there is a very strong correlation between performance and memorization in ICL when it outperforms zero-shot learning. Overall, our study uncovers memorization as a new factor impacting ICL, raising an important question: to what extent do LLMs truly generalize from demonstrations in ICL, and how much of their success is due to memorization?

Alice Saebom Kwak, Jacob O. Israelsen, Clayton T. Morrison et al.

This work introduces a natural language inference (NLI) dataset that focuses on the validity of statements in legal wills. This dataset is unique because: (a) each entailment decision requires three inputs: the statement from the will, the law, and the conditions that hold at the time of the testator's death; and (b) the included texts are longer than the ones in current NLI datasets. We trained eight neural NLI models in this dataset. All the models achieve more than 80% macro F1 and accuracy, which indicates that neural approaches can handle this task reasonably well. However, group accuracy, a stricter evaluation measure that is calculated with a group of positive and negative examples generated from the same statement as a unit, is in mid 80s at best, which suggests that the models' understanding of the task remains superficial. Further ablative analyses and explanation experiments indicate that all three text segments are used for prediction, but some decisions rely on semantically irrelevant tokens. This indicates that overfitting on these longer texts likely happens, and that additional research is required for this task to be solved.

Mohaddeseh Bastan, Nishant Shankar, Mihai Surdeanu et al.

Can language models read biomedical texts and explain the biomedical mechanisms discussed? In this work we introduce a biomedical mechanism summarization task. Biomedical studies often investigate the mechanisms behind how one entity (e.g., a protein or a chemical) affects another in a biological context. The abstracts of these publications often include a focused set of sentences that present relevant supporting statements regarding such relationships, associated experimental evidence, and a concluding sentence that summarizes the mechanism underlying the relationship. We leverage this structure and create a summarization task, where the input is a collection of sentences and the main entities in an abstract, and the output includes the relationship and a sentence that summarizes the mechanism. Using a small amount of manually labeled mechanism sentences, we train a mechanism sentence classifier to filter a large biomedical abstract collection and create a summarization dataset with 22k instances. We also introduce conclusion sentence generation as a pretraining task with 611k instances. We benchmark the performance of large bio-domain language models. We find that while the pretraining task help improves performance, the best model produces acceptable mechanism outputs in only 32% of the instances, which shows the task presents significant challenges in biomedical language understanding and summarization.

Fan Luo, Mihai Surdeanu

Lexical and semantic matches are commonly used as relevance measurements for information retrieval. Together they estimate the semantic equivalence between the query and the candidates. However, semantic equivalence is not the only relevance signal that needs to be considered when retrieving evidences for multi-hop questions. In this work, we demonstrate that textual entailment relation is another important relevance dimension that should be considered. To retrieve evidences that are either semantically equivalent to or entailed by the question simultaneously, we divide the task of evidence retrieval for multi-hop question answering (QA) into two sub-tasks, i.e., semantic textual similarity and inference similarity retrieval. We propose two ensemble models, EAR and EARnest, which tackle each of the sub-tasks separately and then jointly re-rank sentences with the consideration of the diverse relevance signals. Experimental results on HotpotQA verify that our models not only significantly outperform all the single retrieval models it is based on, but is also more effective than two intuitive ensemble baseline models.

Enfa George, Mihai Surdeanu

We introduce SexTok, a multi-modal dataset composed of TikTok videos labeled as sexually suggestive (from the annotator's point of view), sex-educational content, or neither. Such a dataset is necessary to address the challenge of distinguishing between sexually suggestive content and virtual sex education videos on TikTok. Children's exposure to sexually suggestive videos has been shown to have adversarial effects on their development. Meanwhile, virtual sex education, especially on subjects that are more relevant to the LGBTQIA+ community, is very valuable. The platform's current system removes or penalizes some of both types of videos, even though they serve different purposes. Our dataset contains video URLs, and it is also audio transcribed. To validate its importance, we explore two transformer-based models for classifying the videos. Our preliminary results suggest that the task of distinguishing between these types of videos is learnable but challenging. These experiments suggest that this dataset is meaningful and invites further study on the subject.

Zhengzhong Liang, Zeyu Zhang, Steven Bethard et al.

Languages models have been successfully applied to a variety of reasoning tasks in NLP, yet the language models still suffer from compositional generalization. In this paper we present Explainable Verbal Reasoner Plus (EVR+), a reasoning framework that enhances language models' compositional reasoning ability by (1) allowing the model to explicitly generate and execute symbolic operators, and (2) allowing the model to decompose a complex task into several simpler ones in a flexible manner. Compared with its predecessor Explainable Verbal Reasoner (EVR) and other previous approaches adopting similar ideas, our framework supports more diverse types of reasoning such as nested loops and different types of recursion. To evaluate our reasoning framework, we build a synthetic dataset with five tasks that require compositional reasoning. Results show that our reasoning framework can enhance the language model's compositional generalization performance on the five tasks, using a fine-tuned language model. We also discussed the possibility and the challenges to combine our reasoning framework with a few-shot prompted language model.

Sushma Anand Akoju, Robert Vacareanu, Haris Riaz et al.

We introduce a synthetic dataset called Sentences Involving Complex Compositional Knowledge (SICCK) and a novel analysis that investigates the performance of Natural Language Inference (NLI) models to understand compositionality in logic. We produce 1,304 sentence pairs by modifying 15 examples from the SICK dataset (Marelli et al., 2014). To this end, we modify the original texts using a set of phrases - modifiers that correspond to universal quantifiers, existential quantifiers, negation, and other concept modifiers in Natural Logic (NL) (MacCartney, 2009). We use these phrases to modify the subject, verb, and object parts of the premise and hypothesis. Lastly, we annotate these modified texts with the corresponding entailment labels following NL rules. We conduct a preliminary verification of how well the change in the structural and semantic composition is captured by neural NLI models, in both zero-shot and fine-tuned scenarios. We found that the performance of NLI models under the zero-shot setting is poor, especially for modified sentences with negation and existential quantifiers. After fine-tuning this dataset, we observe that models continue to perform poorly over negation, existential and universal modifiers.

Fan Luo, Mihai Surdeanu

Building a question answering (QA) model with less annotation costs can be achieved by utilizing active learning (AL) training strategy. It selects the most informative unlabeled training data to update the model effectively. Acquisition functions for AL are used to determine how informative each training example is, such as uncertainty or diversity based sampling. In this work, we propose a perturbation-based active learning acquisition strategy and demonstrate it is more effective than existing commonly used strategies.

Enrique Noriega-Atala, Mihai Surdeanu, Clayton T. Morrison

We propose a method to teach an automated agent to learn how to search for multi-hop paths of relations between entities in an open domain. The method learns a policy for directing existing information retrieval and machine reading resources to focus on relevant regions of a corpus. The approach formulates the learning problem as a Markov decision process with a state representation that encodes the dynamics of the search process and a reward structure that minimizes the number of documents that must be processed while still finding multi-hop paths. We implement the method in an actor-critic reinforcement learning algorithm and evaluate it on a dataset of search problems derived from a subset of English Wikipedia. The algorithm finds a family of policies that succeeds in extracting the desired information while processing fewer documents compared to several baseline heuristic algorithms.

Razvan-Gabriel Dumitru, Paul-Ioan Clotan, Vikas Yadav et al.

This paper introduces a novel model compression approach through dynamic layer-specific pruning in Large Language Models (LLMs), enhancing the traditional methodology established by SliceGPT. By transitioning from constant to dynamic slicing, our method leverages the newly proposed Layer Redundancy (LR) score, which assesses how much change each layer changes its input by measuring the cosine similarity of the input to the output of the layer. We use this score to prune parts of individual layers based on redundancy in such a way that the average pruned percentage for all layers is a fixed value. We conducted extensive experiments using models like Llama3-8B and Mistral-7B on multiple datasets, evaluating different slicing bases and percentages to determine optimal configurations that balance efficiency and performance. Our findings show that our dynamic slicing approach not only maintains but, in many cases, enhances model performance compared to the baseline established by constant slicing methods. For instance, in several settings, we see performance improvements of up to 5% over the SliceGPT baseline. Additionally, a perplexity decrease by as much as 7% was observed across multiple benchmarks, validating the effectiveness of our method. The code, model weights, and datasets are open-sourced at https://github.com/RazvanDu/DynamicSlicing.

Razvan-Gabriel Dumitru, Minglai Yang, Vikas Yadav et al.

We introduce CopySpec, a simple yet effective technique to tackle the inefficiencies LLMs face when generating responses that closely resemble previous outputs or responses that can be verbatim extracted from context. CopySpec identifies repeated sequences in the model's chat history or context and speculates that the same tokens will follow, enabling seamless copying without compromising output quality and without requiring additional GPU memory. To evaluate the effectiveness of our approach, we conducted experiments using seven LLMs and five datasets: MT-Bench, CNN/DM, GSM8K, HumanEval, and our newly created dataset, MT-Redundant. MT-Redundant, introduced in this paper, transforms the second turn of MT-Bench into a request for variations of the first turn's answer, simulating real-world scenarios where users request modifications to prior responses. Our results demonstrate significant speed-ups: up to 2.35x on CNN/DM, 3.08x on the second turn of select MT-Redundant categories, and 2.66x on the third turn of GSM8K's self-correction tasks. Importantly, we show that CopySpec integrates seamlessly with speculative decoding, yielding an average 49% additional speed-up over speculative decoding for the second turn of MT-Redundant across all eight categories. While LLMs, even with speculative decoding, suffer from slower inference as context size grows, CopySpec leverages larger contexts to accelerate inference, making it a faster complementary solution. Our code and dataset are publicly available at https://github.com/RazvanDu/CopySpec.

Razvan-Gabriel Dumitru, Darius Peteleaza, Mihai Surdeanu

We introduce the concept of multiple temporal perspectives, a novel approach applicable to Recurrent Neural Network (RNN) architectures for enhancing their understanding of sequential data. This method involves maintaining diverse temporal views of previously encountered text, significantly enriching the language models' capacity to interpret context. To show the efficacy of this approach, we incorporate it into the Receptance Weighted Key Value (RWKV) architecture, addressing its inherent challenge of retaining all historical information within a single hidden state. Notably, this improvement is achieved with a minimal increase in the number of parameters --even as little as $0.04\%$ of the original number of parameters. Further, the additional parameters necessary for the multiple temporal perspectives are fine-tuned with minimal computational overhead, avoiding the need for a full pre-training. The resulting model maintains linear computational complexity during prompt inference, ensuring consistent efficiency across various sequence lengths. The empirical results and ablation studies included in our research validate the effectiveness of our approach, showcasing improved performance across multiple benchmarks. The code, model weights and datasets are open-sourced at: https://github.com/RazvanDu/TemporalRNNs.

Md Asiful Islam, Mihai Surdeanu

We propose a lightweight explainable guardrail (LEG) method to detect unsafe prompts. LEG uses a multi-task learning architecture to jointly learn a prompt classifier and an explanation classifier, where the latter labels prompt words that explain the safe/unsafe overall decision. LEG is trained on synthetic explanation data, which is generated using a novel strategy that counteracts the confirmation biases of LLMs. Lastly, LEG's training process uses a novel loss that captures global explanation signals as a weak supervision and combines cross-entropy and focal losses with uncertainty-based weighting. LEG obtains equivalent or better performance than the state-of-the-art for both prompt classification and explainability, both in-domain and out-of-domain on three datasets, despite the fact that its model size is considerably smaller than current approaches.

Razvan-Gabriel Dumitru, Vikas Yadav, Rishabh Maheshwary et al.

We present a simple meta quantization approach that quantizes different layers of a large language model (LLM) at different bit levels, and is independent of the underlying quantization technique. Specifically, we quantize the most important layers to higher bit precision and less important layers to lower bits. We propose two effective strategies to measure the importance of layers within LLMs: the first measures the importance of a layer based on how different its output embeddings are from the input embeddings (higher is better); the second estimates the importance of a layer using the number of layer weights that are much larger than average (smaller is better). We show that quantizing different layers at varying bits according to our importance scores results in minimal performance drop with a far more compressed model size. Finally, we present several practical key takeaways from our variable layer-wise quantization experiments: (a) LLM performance under variable quantization remains close to the original model until 25-50% of layers are moved in lower quantization using our proposed ordering but only until 5-10% if moved using no specific ordering; (b) Adding layer importance to inherently dynamic quantization techniques can further improve their performance, showing that our approach is complementary to other dynamic quantization methods; (c) Quantizing LLMs to lower bits performs substantially better than pruning unless extreme quantization (2-bit) is used; and (d) Layer-wise quantization to lower bits works better in the case of larger LLMs with more layers compared to smaller LLMs with fewer layers. Our code is publicly available at https://github.com/RazvanDu/LayerwiseQuant/.

Haris Riaz, Razvan-Gabriel Dumitru, Mihai Surdeanu

In this work, we revisit the problem of semi-supervised named entity recognition (NER) focusing on extremely light supervision, consisting of a lexicon containing only 10 examples per class. We introduce ELLEN, a simple, fully modular, neuro-symbolic method that blends fine-tuned language models with linguistic rules. These rules include insights such as ''One Sense Per Discourse'', using a Masked Language Model as an unsupervised NER, leveraging part-of-speech tags to identify and eliminate unlabeled entities as false negatives, and other intuitions about classifier confidence scores in local and global context. ELLEN achieves very strong performance on the CoNLL-2003 dataset when using the minimal supervision from the lexicon above. It also outperforms most existing (and considerably more complex) semi-supervised NER methods under the same supervision settings commonly used in the literature (i.e., 5% of the training data). Further, we evaluate our CoNLL-2003 model in a zero-shot scenario on WNUT-17 where we find that it outperforms GPT-3.5 and achieves comparable performance to GPT-4. In a zero-shot setting, ELLEN also achieves over 75% of the performance of a strong, fully supervised model trained on gold data. Our code is available at: https://github.com/hriaz17/ELLEN.

Roya Kabiri, Simin Karimi, Mihai Surdeanu

This paper presents the phonological, morphological, and syntactic distinctions between formal and informal Persian, showing that these two variants have fundamental differences that cannot be attributed solely to pronunciation discrepancies. Given that informal Persian exhibits particular characteristics, any computational model trained on formal Persian is unlikely to transfer well to informal Persian, necessitating the creation of dedicated treebanks for this variety. We thus detail the development of the open-source Informal Persian Universal Dependency Treebank, a new treebank annotated within the Universal Dependencies scheme. We then investigate the parsing of informal Persian by training two dependency parsers on existing formal treebanks and evaluating them on out-of-domain data, i.e. the development set of our informal treebank. Our results show that parsers experience a substantial performance drop when we move across the two domains, as they face more unknown tokens and structures and fail to generalize well. Furthermore, the dependency relations whose performance deteriorates the most represent the unique properties of the informal variant. The ultimate goal of this study that demonstrates a broader impact is to provide a stepping-stone to reveal the significance of informal variants of languages, which have been widely overlooked in natural language processing tools across languages.

Minglai Yang, Xinyu Guo, Mihai Surdeanu et al.

Large Language Models (LLMs) encode factual knowledge within hidden parametric spaces that are difficult to inspect or control. While Sparse Autoencoders (SAEs) can decompose hidden activations into more fine-grained, interpretable features, they often struggle to reliably align these features with human-defined concepts, resulting in entangled and distributed feature representations. To address this, we introduce AlignSAE, a method that aligns SAE features with a defined ontology through a "pre-train, then post-train" curriculum. After an initial unsupervised training phase, we apply supervised post-training to bind specific concepts to dedicated latent slots while preserving the remaining capacity for general reconstruction. This separation creates an interpretable interface where specific relations can be inspected and controlled without interference from unrelated features. Empirical results demonstrate that AlignSAE enables precise causal interventions, such as reliable "concept swaps", by targeting single, semantically aligned slots.

Robert Vacareanu, Vlad-Andrei Negru, Vasile Suciu et al.

We analyze how well pre-trained large language models (e.g., Llama2, GPT-4, Claude 3, etc) can do linear and non-linear regression when given in-context examples, without any additional training or gradient updates. Our findings reveal that several large language models (e.g., GPT-4, Claude 3) are able to perform regression tasks with a performance rivaling (or even outperforming) that of traditional supervised methods such as Random Forest, Bagging, or Gradient Boosting. For example, on the challenging Friedman #2 regression dataset, Claude 3 outperforms many supervised methods such as AdaBoost, SVM, Random Forest, KNN, or Gradient Boosting. We then investigate how well the performance of large language models scales with the number of in-context exemplars. We borrow from the notion of regret from online learning and empirically show that LLMs are capable of obtaining a sub-linear regret.

Fahmida Alam, Mihai Surdeanu, Ellen Riloff

Large language models (LLMs) struggle with relation completion (RC), both with and without retrieval-augmented generation (RAG), particularly when the required information is rare or sparsely represented. To address this, we propose a novel multi-stage paraphrase-guided relation-completion framework, RC-RAG, that systematically incorporates relation paraphrases across multiple stages. In particular, RC-RAG: (a) integrates paraphrases into retrieval to expand lexical coverage of the relation, (b) uses paraphrases to generate relation-aware summaries, and (c) leverages paraphrases during generation to guide reasoning for relation completion. Importantly, our method does not require any model fine-tuning. Experiments with five LLMs on two benchmark datasets show that RC-RAG consistently outperforms several RAG baselines. In long-tail settings, the best-performing LLM augmented with RC-RAG improves by 40.6 Exact Match (EM) points over its standalone performance and surpasses two strong RAG baselines by 16.0 and 13.8 EM points, respectively, while maintaining low computational overhead.

Aunabil Chakma, Mihai Surdeanu, Eduardo Blanco

This paper presents several strategies to automatically obtain additional examples for in-context learning of one-shot relation extraction. Specifically, we introduce a novel strategy for example selection, in which new examples are selected based on the similarity of their underlying syntactic-semantic structure to the provided one-shot example. We show that this method results in complementary word choices and sentence structures when compared to LLM-generated examples. When these strategies are combined, the resulting hybrid system achieves a more holistic picture of the relations of interest than either method alone. Our framework transfers well across datasets (FS-TACRED and FS-FewRel) and LLM families (Qwen and Gemma). Overall, our hybrid selection method consistently outperforms alternative strategies and achieves state-of-the-art performance on FS-TACRED and strong gains on a customized FewRel subset.

Minglai Yang, Ethan Huang, Liang Zhang et al.

We introduce Grade School Math with Distracting Context (GSM-DC), a synthetic benchmark to evaluate Large Language Models' (LLMs) reasoning robustness against systematically controlled irrelevant context (IC). GSM-DC constructs symbolic reasoning graphs with precise distractor injections, enabling rigorous, reproducible evaluation. Our experiments demonstrate that LLMs are significantly sensitive to IC, affecting both reasoning path selection and arithmetic accuracy. Additionally, training models with strong distractors improves performance in both in-distribution and out-of-distribution scenarios. We further propose a stepwise tree search guided by a process reward model, which notably enhances robustness in out-of-distribution conditions.

Robert Vacareanu, Fahmida Alam, Md Asiful Islam et al.

This paper introduces a novel neuro-symbolic architecture for relation classification (RC) that combines rule-based methods with contemporary deep learning techniques. This approach capitalizes on the strengths of both paradigms: the adaptability of rule-based systems and the generalization power of neural networks. Our architecture consists of two components: a declarative rule-based model for transparent classification and a neural component to enhance rule generalizability through semantic text matching. Notably, our semantic matcher is trained in an unsupervised domain-agnostic way, solely with synthetic data. Further, these components are loosely coupled, allowing for rule modifications without retraining the semantic matcher. In our evaluation, we focused on two few-shot relation classification datasets: Few-Shot TACRED and a Few-Shot version of NYT29. We show that our proposed method outperforms previous state-of-the-art models in three out of four settings, despite not seeing any human-annotated training data. Further, we show that our approach remains modular and pliable, i.e., the corresponding rules can be locally modified to improve the overall model. Human interventions to the rules for the TACRED relation \texttt{org:parents} boost the performance on that relation by as much as 26\% relative improvement, without negatively impacting the other relations, and without retraining the semantic matching component.

Fahmida Alam, Md Asiful Islam, Robert Vacareanu et al.

We introduce a meta dataset for few-shot relation extraction, which includes two datasets derived from existing supervised relation extraction datasets NYT29 (Takanobu et al., 2019; Nayak and Ng, 2020) and WIKIDATA (Sorokin and Gurevych, 2017) as well as a few-shot form of the TACRED dataset (Sabo et al., 2021). Importantly, all these few-shot datasets were generated under realistic assumptions such as: the test relations are different from any relations a model might have seen before, limited training data, and a preponderance of candidate relation mentions that do not correspond to any of the relations of interest. Using this large resource, we conduct a comprehensive evaluation of six recent few-shot relation extraction methods, and observe that no method comes out as a clear winner. Further, the overall performance on this task is low, indicating substantial need for future research. We release all versions of the data, i.e., both supervised and few-shot, for future research.

Vlad Andrei Negru, Robert Vacareanu, Camelia Lemnaru et al.

We introduce MorphNLI, a modular step-by-step approach to natural language inference (NLI). When classifying the premise-hypothesis pairs into {entailment, contradiction, neutral}, we use a language model to generate the necessary edits to incrementally transform (i.e., morph) the premise into the hypothesis. Then, using an off-the-shelf NLI model we track how the entailment progresses with these atomic changes, aggregating these intermediate labels into a final output. We demonstrate the advantages of our proposed method particularly in realistic cross-domain settings, where our method always outperforms strong baselines with improvements up to 12.6% (relative). Further, our proposed approach is explainable as the atomic edits can be used to understand the overall NLI label.

Alice Saebom Kwak, Maria Alexeeva, Gus Hahn-Powell et al.

The current trend in information extraction (IE) is to rely extensively on large language models, effectively discarding decades of experience in building symbolic or statistical IE systems. This paper compares a neuro-symbolic (NS) and an LLM-based IE system in the agricultural domain, evaluating them on nine interviews across pork, dairy, and crop subdomains. The LLM-based system outperforms the NS one (F1 total: 69.4 vs. 52.7; core: 63.0 vs. 47.2), where total includes all extracted information and core focuses on essential details. However, each system has trade-offs: the NS approach offers faster runtime, greater control, and high accuracy in context-free tasks but lacks generalizability, struggles with contextual nuances, and requires significant resources to develop and maintain. The LLM-based system achieves higher performance, faster deployment, and easier maintenance but has slower runtime, limited control, model dependency and hallucination risks. Our findings highlight the "hidden cost" of deploying NLP systems in real-world applications, emphasizing the need to balance performance, efficiency, and control.

Xinyu Guo, Zhengliang Shi, Minglai Yang et al.

This paper introduces a framework for relation extraction (RE) that enhances both accuracy and explainability. The framework has two key components: (i) a reasoning mechanism that formulates relation extraction as a series of text-processing steps inspired by cognitive science, and (ii) an optimization process driven by reinforcement learning (RL) with a novel reward function designed to improve both task accuracy and explanation quality. We call our approach CogRE. Our framework addresses the lack of supervision for language-based explanations in traditional RE by promoting outputs that include important relation keywords. These keywords are drawn from a high-quality dictionary that is automatically constructed using an LLM. We evaluate our approach for the task of one-shot RE using two LLMs and two RE datasets. Our experiments show that CogRE improves explanation quality by addressing two common failure patterns in one-shot RE: poor attention focus and limited one-shot learning capability. For example, our cognitive-structured reasoning with Qwen2.5-15B-Instruct on One-shot NYT29 achieves 24.65% F1, surpassing prior reasoning-based designs. Optimizing this approach with RL using our reward further improves performance by +23.46% (absolute). Finally, human evaluation shows that our best model generates relational keywords closely aligned with gold labels, increasing human explanation quality ratings by 54% (relative).

Reza Sanayei, Srdjan Vesic, Eduardo Blanco et al.

Large Language Models (LLMs) excel at linear reasoning tasks but remain underexplored on non-linear structures such as those found in natural debates, which are best expressed as argument graphs. We evaluate whether LLMs can approximate structured reasoning from Computational Argumentation Theory (CAT). Specifically, we use Quantitative Argumentation Debate (QuAD) semantics, which assigns acceptability scores to arguments based on their attack and support relations. Given only dialogue-formatted debates from two NoDE datasets, models are prompted to rank arguments without access to the underlying graph. We test several LLMs under advanced instruction strategies, including Chain-of-Thought and In-Context Learning. While models show moderate alignment with QuAD rankings, performance degrades with longer inputs or disrupted discourse flow. Advanced prompting helps mitigate these effects by reducing biases related to argument length and position. Our findings highlight both the promise and limitations of LLMs in modeling formal argumentation semantics and motivate future work on graph-aware reasoning.

Shahriar Golchin, Yanfei Chen, Rujun Han et al.

Long-context large language models (LLMs) are able to process inputs containing up to several million tokens. In the scope of in-context learning (ICL), this translates into using hundreds/thousands of demonstrations in the input prompt, enabling many-shot ICL. In practice, a fixed set of demonstrations is often selected at random in many-shot settings due to (1) high inference costs, (2) the benefits of caching and reusing computations, and (3) the similar performance offered by this strategy compared to others when scaled. In this work, we propose two straightforward strategies for demonstration selection in many-shot ICL that improve performance with minimal computational overhead. Our first method combines a small number of demonstrations, selected based on their similarity to each test sample, with a disproportionately larger set of random demonstrations that are cached. The second strategy improves the first by replacing random demonstrations with those selected using centroids derived from test sample representations via k-means clustering. Our experiments with Gemini Pro and Flash across several datasets indicate that our strategies consistently outperform random selection and surpass or match the most performant selection approach while supporting caching and reducing inference cost by up to an order of magnitude. We also show that adjusting the proportion of demonstrations selected based on different criteria can balance performance and inference cost in many-shot ICL.

Enfa Fane, Mihai Surdeanu, Eduardo Blanco et al.

Understanding how news narratives frame entities is crucial for studying media's impact on societal perceptions of events. In this paper, we evaluate the zero-shot capabilities of large language models (LLMs) in classifying framing roles. Through systematic experimentation, we assess the effects of input context, prompting strategies, and task decomposition. Our findings show that a hierarchical approach of first identifying broad roles and then fine-grained roles, outperforms single-step classification. We also demonstrate that optimal input contexts and prompts vary across task levels, highlighting the need for subtask-specific strategies. We achieve a Main Role Accuracy of 89.4% and an Exact Match Ratio of 34.5%, demonstrating the effectiveness of our approach. Our findings emphasize the importance of tailored prompt design and input context optimization for improving LLM performance in entity framing.

Haris Riaz, Ellen Riloff, Mihai Surdeanu

We propose a simple, unsupervised method that injects pragmatic principles in retrieval-augmented generation (RAG) frameworks such as Dense Passage Retrieval to enhance the utility of retrieved contexts. Our approach first identifies which sentences in a pool of documents retrieved by RAG are most relevant to the question at hand, cover all the topics addressed in the input question and no more, and then highlights these sentences within their context, before they are provided to the LLM, without truncating or altering the context in any other way. We show that this simple idea brings consistent improvements in experiments on three question answering tasks (ARC-Challenge, PubHealth and PopQA) using five different LLMs. It notably enhances relative accuracy by up to 19.7% on PubHealth and 10% on ARC-Challenge compared to a conventional RAG system.

Portia Cooper, Harshita Narnoli, Mihai Surdeanu

Text-to-image models are vulnerable to the stepwise "Divide-and-Conquer Attack" (DACA) that utilize a large language model to obfuscate inappropriate content in prompts by wrapping sensitive text in a benign narrative. To mitigate stepwise DACA attacks, we propose a two-layer method involving text summarization followed by binary classification. We assembled the Adversarial Text-to-Image Prompt (ATTIP) dataset ($N=940$), which contained DACA-obfuscated and non-obfuscated prompts. From the ATTIP dataset, we created two summarized versions: one generated by a small encoder model and the other by a large language model. Then, we used an encoder classifier and a GPT-4o classifier to perform content moderation on the summarized and unsummarized prompts. When compared with a classifier that operated over the unsummarized data, our method improved F1 score performance by 31%. Further, the highest recorded F1 score achieved (98%) was produced by the encoder classifier on a summarized ATTIP variant. This study indicates that pre-classification text summarization can inoculate content detection models against stepwise DACA obfuscations.

Robert Vacareanu, Marco A. Valenzuela-Escarcega, George C. G. Barbosa et al.

While deep learning approaches to information extraction have had many successes, they can be difficult to augment or maintain as needs shift. Rule-based methods, on the other hand, can be more easily modified. However, crafting rules requires expertise in linguistics and the domain of interest, making it infeasible for most users. Here we attempt to combine the advantages of these two directions while mitigating their drawbacks. We adapt recent advances from the adjacent field of program synthesis to information extraction, synthesizing rules from provided examples. We use a transformer-based architecture to guide an enumerative search, and show that this reduces the number of steps that need to be explored before a rule is found. Further, we show that without training the synthesis algorithm on the specific domain, our synthesized rules achieve state-of-the-art performance on the 1-shot scenario of a task that focuses on few-shot learning for relation classification, and competitive performance in the 5-shot scenario.

Andrew Zupon, Andrew Carnie, Michael Hammond et al.

Annotation inconsistencies between data sets can cause problems for low-resource NLP, where noisy or inconsistent data cannot be as easily replaced compared with resource-rich languages. In this paper, we propose a method for automatically detecting annotation mismatches between dependency parsing corpora, as well as three related methods for automatically converting the mismatches. All three methods rely on comparing an unseen example in a new corpus with similar examples in an existing corpus. These three methods include a simple lexical replacement using the most frequent tag of the example in the existing corpus, a GloVe embedding-based replacement that considers a wider pool of examples, and a BERT embedding-based replacement that uses contextualized embeddings to provide examples fine-tuned to our specific data. We then evaluate these conversions by retraining two dependency parsers -- Stanza (Qi et al. 2020) and Parsing as Tagging (PaT) (Vacareanu et al. 2020) -- on the converted and unconverted data. We find that applying our conversions yields significantly better performance in many cases. Some differences observed between the two parsers are observed. Stanza has a more complex architecture with a quadratic algorithm, so it takes longer to train, but it can generalize better with less data. The PaT parser has a simpler architecture with a linear algorithm, speeding up training time but requiring more training data to reach comparable or better performance.

Enrique Noriega-Atala, Peter M. Lovett, Clayton T. Morrison et al.

We introduce a family of deep-learning architectures for inter-sentence relation extraction, i.e., relations where the participants are not necessarily in the same sentence. We apply these architectures to an important use case in the biomedical domain: assigning biological context to biochemical events. In this work, biological context is defined as the type of biological system within which the biochemical event is observed. The neural architectures encode and aggregate multiple occurrences of the same candidate context mentions to determine whether it is the correct context for a particular event mention. We propose two broad types of architectures: the first type aggregates multiple instances that correspond to the same candidate context with respect to event mention before emitting a classification; the second type independently classifies each instance and uses the results to vote for the final class, akin to an ensemble approach. Our experiments show that the proposed neural classifiers are competitive and some achieve better performance than previous state of the art traditional machine learning methods without the need for feature engineering. Our analysis shows that the neural methods particularly improve precision compared to traditional machine learning classifiers and also demonstrates how the difficulty of inter-sentence relation extraction increases as the distance between the event and context mentions increase.

Hoang Van, Zheng Tang, Mihai Surdeanu

The general goal of text simplification (TS) is to reduce text complexity for human consumption. This paper investigates another potential use of neural TS: assisting machines performing natural language processing (NLP) tasks. We evaluate the use of neural TS in two ways: simplifying input texts at prediction time and augmenting data to provide machines with additional information during training. We demonstrate that the latter scenario provides positive effects on machine performance on two separate datasets. In particular, the latter use of TS improves the performances of LSTM (1.82-1.98%) and SpanBERT (0.7-1.3%) extractors on TACRED, a complex, large-scale, real-world relation extraction task. Further, the same setting yields improvements of up to 0.65% matched and 0.62% mismatched accuracies for a BERT text classifier on MNLI, a practical natural language inference dataset.

Hoang Van, Vikas Yadav, Mihai Surdeanu

We propose a simple and effective strategy for data augmentation for low-resource machine reading comprehension (MRC). Our approach first pretrains the answer extraction components of a MRC system on the augmented data that contains approximate context of the correct answers, before training it on the exact answer spans. The approximate context helps the QA method components in narrowing the location of the answers. We demonstrate that our simple strategy substantially improves both document retrieval and answer extraction performance by providing larger context of the answers and additional training data. In particular, our method significantly improves the performance of BERT based retriever (15.12\%), and answer extractor (4.33\% F1) on TechQA, a complex, low-resource MRC task. Further, our data augmentation strategy yields significant improvements of up to 3.9\% exact match (EM) and 2.7\% F1 for answer extraction on PolicyQA, another practical but moderate sized QA dataset that also contains long answer spans.

Hoang Van, Ahmad Musa, Mihai Surdeanu et al.

We study the language of food on Twitter during the pandemic lockdown in the United States, focusing on the two month period of March 15 to May 15, 2020. Specifically, we analyze over770,000 tweets published during the lockdown and the equivalent period in the five previous years and highlight several worrying trends. First, we observe that during the lockdown there was a notable shift from mentions of healthy foods to unhealthy foods. Second, we show an increased pointwise mutual information of depression hashtags with food-related tweets posted during the lockdown and an increased association between depression hashtags and unhealthy foods, tobacco, and alcohol during the lockdown.

Zhengzhong Liang, Yiyun Zhao, Mihai Surdeanu

Evidence retrieval is a key component of explainable question answering (QA). We argue that, despite recent progress, transformer network-based approaches such as universal sentence encoder (USE-QA) do not always outperform traditional information retrieval (IR) methods such as BM25 for evidence retrieval for QA. We introduce a lexical probing task that validates this observation: we demonstrate that neural IR methods have the capacity to capture lexical differences between questions and answers, but miss obvious lexical overlap signal. Learning from this probing analysis, we introduce a hybrid approach for evidence retrieval that combines the advantages of both IR directions. Our approach uses a routing classifier that learns when to direct incoming questions to BM25 vs. USE-QA for evidence retrieval using very simple statistics, which can be efficiently extracted from the top candidate evidence sentences produced by a BM25 model. We demonstrate that this hybrid evidence retrieval generally performs better than either individual retrieval strategy on three QA datasets: OpenBookQA, ReQA SQuAD, and ReQA NQ. Furthermore, we show that the proposed routing strategy is considerably faster than neural methods, with a runtime that is up to 5 times faster than USE-QA.

Vikas Yadav, Steven Bethard, Mihai Surdeanu

Evidence retrieval is a critical stage of question answering (QA), necessary not only to improve performance, but also to explain the decisions of the corresponding QA method. We introduce a simple, fast, and unsupervised iterative evidence retrieval method, which relies on three ideas: (a) an unsupervised alignment approach to soft-align questions and answers with justification sentences using only GloVe embeddings, (b) an iterative process that reformulates queries focusing on terms that are not covered by existing justifications, which (c) a stopping criterion that terminates retrieval when the terms in the given question and candidate answers are covered by the retrieved justifications. Despite its simplicity, our approach outperforms all the previous methods (including supervised methods) on the evidence selection task on two datasets: MultiRC and QASC. When these evidence sentences are fed into a RoBERTa answer classification component, we achieve state-of-the-art QA performance on these two datasets.

Vikas Yadav, Steven Bethard, Mihai Surdeanu

We propose an unsupervised strategy for the selection of justification sentences for multi-hop question answering (QA) that (a) maximizes the relevance of the selected sentences, (b) minimizes the overlap between the selected facts, and (c) maximizes the coverage of both question and answer. This unsupervised sentence selection method can be coupled with any supervised QA approach. We show that the sentences selected by our method improve the performance of a state-of-the-art supervised QA model on two multi-hop QA datasets: AI2's Reasoning Challenge (ARC) and Multi-Sentence Reading Comprehension (MultiRC). We obtain new state-of-the-art performance on both datasets among approaches that do not use external resources for training the QA system: 56.82% F1 on ARC (41.24% on Challenge and 64.49% on Easy) and 26.1% EM0 on MultiRC. Our justification sentences have higher quality than the justifications selected by a strong information retrieval baseline, e.g., by 5.4% F1 in MultiRC. We also show that our unsupervised selection of justification sentences is more stable across domains than a state-of-the-art supervised sentence selection method.

Sandeep Suntwal, Mithun Paul, Rebecca Sharp et al.

In this work we aim to understand and estimate the importance that a neural network assigns to various aspects of the data while learning and making predictions. Here we focus on the recognizing textual entailment (RTE) task and its application to fact verification. In this context, the contributions of this work are as follows. We investigate the attention weights a state of the art RTE method assigns to input tokens in the RTE component of fact verification systems, and confirm that most of the weight is assigned to POS tags of nouns (e.g., NN, NNP etc.) or their phrases. To verify that these lexicalized models transfer poorly, we implement a domain transfer experiment where a RTE component is trained on the FEVER data, and tested on the Fake News Challenge (FNC) dataset. As expected, even though this method achieves high accuracy when evaluated in the same domain, the performance in the target domain is poor, marginally above chance.To mitigate this dependence on lexicalized information, we experiment with several strategies for masking out names by replacing them with their semantic category, coupled with a unique identifier to mark that the same or new entities are referenced between claim and evidence. The results show that, while the performance on the FEVER dataset remains at par with that of the model trained on lexicalized data, it improves significantly when tested in the FNC dataset. Thus our experiments demonstrate that our strategy is successful in mitigating the dependency on lexical information.

Vikas Yadav, Rebecca Sharp, Mihai Surdeanu

While increasingly complex approaches to question answering (QA) have been proposed, the true gain of these systems, particularly with respect to their expensive training requirements, can be inflated when they are not compared to adequate baselines. Here we propose an unsupervised, simple, and fast alignment and information retrieval baseline that incorporates two novel contributions: a \textit{one-to-many alignment} between query and document terms and \textit{negative alignment} as a proxy for discriminative information. Our approach not only outperforms all conventional baselines as well as many supervised recurrent neural networks, but also approaches the state of the art for supervised systems on three QA datasets. With only three hyperparameters, we achieve 47\% P@1 on an 8th grade Science QA dataset, 32.9\% P@1 on a Yahoo! answers QA dataset and 64\% MAP on WikiQA. We also achieve 26.56\% and 58.36\% on ARC challenge and easy dataset respectively. In addition to including the additional ARC results in this version of the paper, for the ARC easy set only we also experimented with one additional parameter -- number of justifications retrieved.