Papers by Mihai Surdeanu
The Lies Characters Tell: Utilizing Large Language Models to Normalize Adversarial Unicode Perturbations (2025.findings-acl)
Copied to clipboard
| Challenge: | Homoglyphs are visually homogeneous to Latin letters and are used to mask offensive content. |
| Approach: | They propose two methods to normalize homoglyphs by replacing non-Latin characters with a delimiter and using large language models to determine which characters should be replaced with Latin letters. |
| Outcome: | The proposed methods normalize homoglyphs by replacing non-Latin characters with a delimiter and prompting large language models to "fill in the blanks" the authors found that the proposed methods produced normalized text with an average cosine similarity score of 0.91 to the original tweets and 0.96 to the tweets using the direct method. |
Grounding Gradable Adjectives through Crowdsourcing (L18-1)
Copied to clipboard
| Challenge: | Often, texts describe interactions using vague, high-level language . crowdsourcing is expensive and requires extensive literature review and time . |
| Approach: | They propose a method for estimating concrete groundings for a set of gradable adjectives by crowdsourcing human intuitions and fitting a mixed effects model to the text. |
| Outcome: | The proposed model can generalize to unseen data and has a predictive R 2 of 0.632 in general and 0.677 on a subset of high-frequency adjectives. |
Active Learning Design Choices for NER with Transformers (2024.lrec-main)
Copied to clipboard
Robert Vacareanu, Enrique Noriega-Atala, Gus Hahn-Powell, Marco A. Valenzuela-Escarcega, Mihai Surdeanu
| Challenge: | In the field of natural language processing, active learning is a technique that is used to decide which examples are worth annotating . a number of studies have focused on sequence classification, text classification, question answering, and question answering. |
| Approach: | They propose two different approaches to deal with partially-annotated sentences . they propose an annotation scheme that can be used to train with tokens . |
| Outcome: | The proposed approaches achieve comparable or better performance than sentence-level annotations with a smaller number of annotated tokens. |
If You Want to Go Far Go Together: Unsupervised Joint Candidate Evidence Retrieval for Multi-hop Question Answering (2021.naacl-main)
Copied to clipboard
| Challenge: | et al. : evidence retrieval is highly dependent on partial, incorrect or no supporting knowledge. |
| Approach: | They propose a method that retrieves and reranks evidence facts jointly . they propose to account for links between sentences and coverage with the given query . |
| Outcome: | The proposed approach achieves state-of-the-art evidence retrieval performance on two multi-hop question answering datasets. |
Validity Assessment of Legal Will Statements as Natural Language Inference (2022.findings-emnlp)
Copied to clipboard
| Challenge: | This study introduces a dataset that focuses on the validity of statements in legal wills. |
| Approach: | They propose a dataset that focuses on the validity of statements in legal wills. |
| Outcome: | The proposed model achieves 80% macro F1 and accuracy, but group accuracy is in mid 80s at best, suggesting that the models’ understanding of the task remains superficial. |
Text Annotation Graphs: Annotating Complex Natural Language Phenomena (L18-1)
Copied to clipboard
| Challenge: | Text Annotation Graphs is a web-based tool for annotating text . it provides functionality for representing complex relationships between words and word phrases . |
| Approach: | They introduce a web-based tool for annotating text, Text Annotation Graphs, or TAG . it provides functionality for representing complex relationships between words and word phrases . |
| Outcome: | The proposed software can represent complex relationships between words and words . it can also be used to find similar structures within the current document or external annotated documents. |
Explainable Multi-hop Verbal Reasoning Through Internal Monologue (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing state-of-the-art language models do not make intermediate reasoning steps explicit . large pretrained language models such as BERT and RoBERTa have been successfully used in multi-hop reasoning problems . |
| Approach: | They propose to decompose multi-hop reasoning problems into several simple ones and use natural language to guide intermediate reasoning hops. |
| Outcome: | The proposed model can generate subgoals and perform inference in natural language at each reasoning step. |
Bootstrapping Neural Relation and Explanation Classifiers (2023.acl-short)
Copied to clipboard
| Challenge: | supervised approaches that use only rules to explain the outputs of the relation classifier are data hungry and expensive to obtain. |
| Approach: | They propose a method that self trains (or bootstraps) neural relation and explanation classifiers by iterating the outputs into rules and applying them to unlabeled text to produce new annotations. |
| Outcome: | The proposed method outperforms the rule-based model on the TACRED dataset by 15 F1 points and performs comparatively with the prompt-based approach without an additional natural language inference component. |
Data and Model Distillation as a Solution for Domain-transferable Fact Verification (2021.naacl-main)
Copied to clipboard
| Challenge: | Neural networks depend heavily on lexicalized information, which transfers poorly between domains. |
| Approach: | They propose a method to delexicize lexicalized data and a model distillation technique to prevent aggressive data distillation. |
| Outcome: | The proposed method improves performance on lexicalized data and out of domain models. |
BioNLI: Generating a Biomedical NLI Dataset Using Lexico-semantic Constraints for Adversarial Examples (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Biomedical research has progressed at a tremendous pace, with PubMed2 indexing well over 1M publications per year in the past eight years. |
| Approach: | They propose a semi-supervised procedure that bootstraps biomedical NLI datasets from positive entailment examples present in biomedically published texts. |
| Outcome: | The proposed procedure bootstraps biomedical NLI datasets from positive entailment examples from biomedically challenging texts. |
MorphNLI: A Stepwise Approach to Natural Language Inference Using Text Morphing (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing models fail to capture important semantic features of logic such as monotonicity and negation. |
| Approach: | They propose a modular step-by-step approach to natural language inference . they use a language model to generate edits to incrementally transform the premise into the hypothesis . |
| Outcome: | The proposed method outperforms baseline models in realistic cross-domain settings with improvements up to 12.6% (relative). |
ELLEN: Extremely Lightly Supervised Learning for Efficient Named Entity Recognition (2024.lrec-main)
Copied to clipboard
| Challenge: | Named entity recognition (NER) tasks require an amount of annotations that are unrealistic for many real-world applications. |
| Approach: | They propose a semi-supervised named entity recognition method that blends language models with linguistic rules. |
| Outcome: | The proposed method outperforms most existing semi-supervised methods under the same supervision settings commonly used in the literature. |
It’s not Sexually Suggestive; It’s Educative | Separating Sex Education from Suggestive Content on TikTok videos (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing systems for TikTok remove/punish both types of videos, even though they serve different purposes. |
| Approach: | They propose a dataset that contains TikTok videos labeled as sexually suggestive, sex-educational content, or neither . they explore two transformer-based models for classifying the videos to validate their importance . |
| Outcome: | The proposed dataset is useful and invites further study on the subject. |
Enabling Search and Collaborative Assembly of Causal Interactions Extracted from Multilingual and Multi-domain Free Text (N19-4)
Copied to clipboard
George C. G. Barbosa, Zechy Wong, Gus Hahn-Powell, Dane Bell, Rebecca Sharp, Marco A. Valenzuela-Escárcega, Mihai Surdeanu
| Challenge: | a new searchable knowledge graph allows users to search for causal interactions in multiple languages . a recent study shows that search tools are shallow and do not support multilingual research . |
| Approach: | They propose a system that integrates causal interactions into a single searchable knowledge graph. |
| Outcome: | The proposed system extracts over 600 thousand causal statements from 120 thousand Portuguese publications with a precision of 62%. |
Keep Your Bearings: Lightly-Supervised Information Extraction with Ladder Networks That Avoids Semantic Drift (N18-2)
Copied to clipboard
| Challenge: | Using ladder networks, semi-supervised learning can be iterative and drifts semantically as learning progresses. |
| Approach: | They propose a method that uses ladder networks to perform a task of named entity classification using a large, unannotated dataset. |
| Outcome: | The proposed method improves on two standard datasets for named entity classification. |
Unsupervised Alignment-based Iterative Evidence Retrieval for Multi-hop Question Answering (2020.acl-main)
Copied to clipboard
| Challenge: | Evidence retrieval is a critical stage of question answering (QA) . Several multi-hop QA datasets have been proposed recently . |
| Approach: | They propose an unsupervised method that uses only GloVe embeddings to soft-align questions with justification sentences and an iterative process that reformulates queries focusing on terms that are not covered by existing justifications. |
| Outcome: | The proposed method outperforms all previous methods on the evidence selection task on two datasets: MultiRC and QASC. |
From Examples to Rules: Neural Guided Rule Synthesis for Information Extraction (2022.lrec-1)
Copied to clipboard
Robert Vacareanu, Marco A. Valenzuela-Escárcega, George Caique Gouveia Barbosa, Rebecca Sharp, Gustave Hahn-Powell, Mihai Surdeanu
| Challenge: | a "deep learning tsunami" has brought tremendous improvements in performance to most NLP applications. |
| Approach: | They propose a method for rule synthesis from examples that combines the advantages of deep learning and rule-based methods. |
| Outcome: | The proposed method achieves state-of-the-art on 1-shot task and competitive performance in 5-shot scenario. |
Information Extraction from Legal Wills: How Well Does GPT-4 Do? (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Using information extraction from legal wills is an important application of artificial intelligence (AI) |
| Approach: | They propose a manually annotated dataset for Information Extraction (IE) from legal wills . they also use it to evaluate the performance of large language models (LLMs) |
| Outcome: | The proposed dataset can be used to evaluate large language models on IE from legal wills . it shows that the model performs reasonably well, but inconsistent outputs and overgeneralization are observed . |
Eidos, INDRA, & Delphi: From Free Text to Executable Causal Models (N19-4)
Copied to clipboard
Rebecca Sharp, Adarsh Pyarelal, Benjamin Gyori, Keith Alcock, Egoitz Laparra, Marco A. Valenzuela-Escárcega, Ajay Nagesh, Vikas Yadav, John Bachman, Zheng Tang, Heather Lent, Fan Luo, Mithun Paul, Steven Bethard, Kobus Barnard, Clayton Morrison, Mihai Surdeanu
| Challenge: | a paper proposes a method for building probabilistic models of complex phenomena such as food insecurity . currently, these models are hand-built for each new situation and require months to construct . |
| Approach: | They propose an approach that builds executable probabilistic models from raw, free text. |
| Outcome: | The proposed approach builds executable probabilistic models from raw, free text. |
Visual Supervision in Bootstrapped Information Extraction (D18-1)
Copied to clipboard
| Challenge: | a list-based interface populated with informative samples is effective for data annotation . a 2D scatterplot populated by diverse and representative samples yields improved models . |
| Approach: | They propose a list-based interface that can be used to build efficient and effective data annotation models. |
| Outcome: | The proposed model learns the distributional similarity of entities through the patterns that match them in a large corpus while being discriminative with respect to human-labeled and machine-promoted entities. |
Automatic Correction of Syntactic Dependency Annotation Differences (2022.lrec-1)
Copied to clipboard
| Challenge: | Annotation inconsistencies between data sets can cause problems for low-resource NLP . a simple method for automatically detecting annotation mismatches between corpora is proposed . |
| Approach: | They propose a method for automatically detecting annotation mismatches between dependency parsing corpora . they also propose three related methods for automatically configuring the mismatch . |
| Outcome: | The proposed method improves performance on both converted and unconverted data. |
A Human-machine Interface for Few-shot Rule Synthesis for Information Extraction (2022.naacl-demo)
Copied to clipboard
Robert Vacareanu, George C.G. Barbosa, Enrique Noriega-Atala, Gus Hahn-Powell, Rebecca Sharp, Marco A. Valenzuela-Escárcega, Mihai Surdeanu
| Challenge: | Vacareanu et al., 2021) proposes a system that helps users build transparent information extraction models . rule-based methods address the opacity of neural architectures by producing models that are transparent . |
| Approach: | They propose a system that assists a user in constructing transparent information extraction models . the system generates high-precision rules even in a 1-shot setting, they show . |
| Outcome: | The proposed system generates high-precision rules even in a 1-shot setting . it outperforms manually written patterns on a widely-used relation extraction dataset . |
Do Transformer Networks Improve the Discovery of Rules from Text? (2022.lrec-1)
Copied to clipboard
| Challenge: | Existing rules that model binary relations such as X treat Y are based on the distributional hypothesis of Harris (1954). |
| Approach: | They propose to implement the distributional hypothesis using contextualized embeddings provided by a transformer-network-based language model to measure the similarity between slots instead of lexical overlap. |
| Outcome: | The proposed approach outperforms the original DIRT algorithm in the question answering-based evaluation. |
Exploring Interpretability in Event Extraction: Multitask Learning of a Neural Event Classifier and an Explanation Decoder (2020.acl-srw)
Copied to clipboard
| Challenge: | EE is a key requirement for machine learning in many domains, e.g., legal, medical, finance. |
| Approach: | They propose an interpretable approach for event extraction that jointly trains a classifier and a rule decoder for event processing. |
| Outcome: | The proposed approach can be used for semi-supervised learning and its performance improves when trained on automatically-labeled data generated by a rule-based system. |
Towards Realistic Few-Shot Relation Extraction: A New Meta Dataset and Evaluation (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods for few-shot relation extraction are not realistic due to the large amount of training data required. |
| Approach: | They propose a meta dataset for few-shot relation extraction based on existing supervised relation extraction datasets and a few-shot form of the TACRED dataset. |
| Outcome: | The proposed methods perform poorly on the few-shot relation extraction task. |
Towards the Necessity for Debiasing Natural Language Inference Datasets (2020.lrec-1)
Copied to clipboard
| Challenge: | Delexicalization of datasets can improve natural language inference performance . a dataset with a delexicalized version of the FEVER dataset is used for natural language learning . |
| Approach: | They propose two techniques for delexicalization that modify annotated datasets to control the importance of lexical entities. |
| Outcome: | The proposed methods maintain performance in-domain and improve performance in some out-of-domain settings. |
Parsing as Tagging (2020.lrec-1)
Copied to clipboard
| Challenge: | Existing methods for dependency parsing treat parse as tagging, but they are not perfect. |
| Approach: | They propose a simple yet accurate method that treats parsing as tagging . they use a sequence model with a bidirectional LSTM over BERT embeddings . |
| Outcome: | The proposed method outperforms the state-of-the-art method on universal dependency (UD) by 1.76% unlabeled attachment score (UAS) for English, 1.98% UAS for French, and 1.16% UAS in German. |
An Exploration of Three Lightly-supervised Representation Learning Approaches for Named Entity Classification (C18-1)
Copied to clipboard
| Challenge: | a recent study compares semi-supervised learning methods with bootstrapping methods . semi-semi-supervised methods reduce the amount of semantic drift introduced by iterative approaches . |
| Approach: | They propose to adapt three semi-supervised representation learning methods to an information extraction task . they show that all methods outperform state-of-the-art semi-representation learning methods . |
| Outcome: | The proposed methods outperform state-of-the-art semi-supervised methods on named entity classification task. |
Informal Persian Universal Dependency Treebank (2022.lrec-1)
Copied to clipboard
| Challenge: | phonological, morphological, and syntactic distinctions between formal and informal Persian are important . formal Persian is not a universally recognized form of language, but is a dialect of informal Persian . |
| Approach: | They develop an open-source treebank for informal Persian to train dependency parsers . they then train dependency lexicographers on existing treebanks and evaluate them on out-of-domain data . |
| Outcome: | The proposed treebanks show that they perform poorly when training on formal and informal Persians. |
Structured Semantic Information Helps Retrieve Better Examples for In-Context Learning Applied to Few-Shot Relation Extraction (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to relation extraction require many training examples per relation, resulting in low results. |
| Approach: | They propose a strategy where new examples are selected based on their similarity to the provided 1-shot example. |
| Outcome: | The proposed strategy outperforms other methods on FS-TACRED and FS - FewRel subsets and achieves state-of-the-art performance on both datasets. |
What does the language of foods say about us? (D19-62)
Copied to clipboard
| Challenge: | Using a dataset of 24 million food-related tweets, we can predict if states in the United States are above the median rates for type 2 diabetes mellitus (T2DM) income, poverty, and education are important factors in predicting T2DM rates, but socioeconomic factors do not capture this information. |
| Approach: | They use a dataset of 24 million food-related tweets to investigate the signal contained in the language of food on social media. |
| Outcome: | The language of food can predict health risks, political orientation, and geographic location, and outperform previous work by 4–18%. |
On the Importance of Delexicalization for Fact Verification (D19-1)
Copied to clipboard
| Challenge: | Neural networks (NNs) perform state-of-the-art (SOA) performance in many complex tasks. |
| Approach: | They investigate the importance that a model assigns to various aspects of data . they experiment with two strategies of masking to mitigate this dependence on lexicalized information . |
| Outcome: | The proposed model improves on the in-domain dataset by 10% compared to the fully lexicalized model. |
CopySpec: Accelerating LLMs with Speculative Copy-and-Paste (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) suffer from slower inference as context size grows, but CopySpec leverages larger contexts to accelerate inference. |
| Approach: | They propose a technique that speculates that the same tokens will follow repeated sequences in the model’s chat history or context and enables seamless copying without compromising output quality. |
| Outcome: | The proposed technique can generate responses that closely resemble previous outputs or responses that can be verbatim extracted from context without compromising output quality and without requiring additional GPU memory. |
Best of Both Worlds: A Pliable and Generalizable Neuro-Symbolic Approach for Relation Classification (2024.findings-naacl)
Copied to clipboard
| Challenge: | a novel neuro-symbolic architecture for relation classification combines rule-based methods with deep learning techniques. |
| Approach: | They propose a neuro-symbolic architecture for relation classification that combines rule-based methods with deep learning techniques. |
| Outcome: | The proposed approach outperforms state-of-the-art models in three out of four settings . human interventions boost the performance on the relation org:parents by as much as 26% relative improvement . |
Change Is the Only Constant: Dynamic LLM Slicing based on Layer Redundancy (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Using dynamic slicing, large language models can be used to reduce computational burden and improve performance. |
| Approach: | They propose a dynamic layer-specific pruning approach that leverages the newly proposed Layer Redundancy score to prune parts of individual layers based on redundancy. |
| Outcome: | The proposed method maintains and enhances model performance over the SliceGPT baseline. |
A Lightweight Explainable Guardrail for Prompt Safety (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods to detect unsafe prompts are based on reinforcement learning with human feedback (RLHF) or direct preference optimization (DPO) . however, these methods lack explainability and are costly to implement. |
| Approach: | They propose a lightweight explainable guardrail method to detect unsafe prompts using a multi-task learning architecture and a novel strategy to counteract confirmation biases. |
| Outcome: | The proposed method obtains equivalent or better performance than the state-of-the-art for both prompt classification and explainability on three datasets. |
Variable Layerwise Quantization: A Simple and Effective Approach to Quantize LLMs (2025.findings-acl)
Copied to clipboard
Razvan-Gabriel Dumitru, Vikas Yadav, Rishabh Maheshwary, Paul Ioan Clotan, Sathwik Tejaswi Madhusudhan, Mihai Surdeanu
| Challenge: | a meta quantization approach quantizes different layers of a large language model at different bit levels. |
| Approach: | They propose a meta quantization approach that quantizes different layers of a large language model at different bit levels. |
| Outcome: | The proposed method quantizes the most important layers to higher bit precision and less important layers at lower bits. |
Quick and (not so) Dirty: Unsupervised Selection of Justification Sentences for Multi-hop Question Answering (D19-1)
Copied to clipboard
| Challenge: | Arras et al., 2017) suggest an unsupervised strategy for the selection of justification sentences for multi-hop question answering (QA) . |
| Approach: | They propose an unsupervised strategy for the selection of justification sentences for multi-hop question answering that maximizes the relevance of the selected sentences, minimizes overlap between selected facts, and maximizes coverage of both question and answer. |
| Outcome: | The proposed strategy improves state-of-the-art supervised QA model on two multi-hop QA datasets: AI2’s Reasoning Challenge (ARC) and Multi-Sentence Reading Comprehension (MultiRC). |
How Is LLM Reasoning Distracted by Irrelevant Context? An Analysis Using a Controlled Benchmark (2025.emnlp-main)
Copied to clipboard
| Challenge: | Prior work has not explored the mechanisms underlying this sensitivity. |
| Approach: | They propose a synthetic benchmark to evaluate Large Language Models’ reasoning robustness against systematically controlled irrelevant context (IC). |
| Outcome: | The proposed model improves in-distribution and out-of-disttribution scenarios while training with strong distractors. |
When and Where Did it Happen? An Encoder-Decoder Model to Identify Scenario Context (2024.findings-emnlp)
Copied to clipboard
Enrique Noriega-Atala, Robert Vacareanu, Salena Ashton, Adarsh Pyarelal, Clayton Morrison, Mihai Surdeanu
| Challenge: | a relatively small fine-tuned encoder-decoder model performs better than out-of-the-box LLMs and semantic role labeling parsers to accurately predict the relevant scenario information. |
| Approach: | They propose a neural architecture finetuned for the task of scenario context generation . they use a curated dataset of time and location annotations to train an encoder-decoder architecture . |
| Outcome: | The proposed model performs better than out-of-the-box LLMs and semantic role labeling parsers to accurately predict the relevant scenario information of a particular entity or event. |
SuMe: A Dataset Towards Summarizing Biomedical Mechanisms (2022.lrec-1)
Copied to clipboard
| Challenge: | Biomedical studies often examine how one entity affects another in a biological context. |
| Approach: | They propose a biomedical mechanism summarization task that pairs biomedically relevant texts with their summaries. |
| Outcome: | The proposed task improves performance but produces acceptable outputs in 32% of instances. |
NEUROSTRUCTURAL DECODING: Neural Text Generation with Structural Constraints (2023.acl-long)
Copied to clipboard
| Challenge: | Current approaches for conditional text generation focus on lexical constraints, but lack syntactic constraints to support complex semantic constraints. |
| Approach: | They propose a decoding algorithm that incorporates syntactic constraints to improve the quality of the generated text. |
| Outcome: | The proposed method improves on three different language generation tasks and shows improved lexical and syntactic metrics. |
How May I Help You? Using Neural Text Simplification to Improve Downstream NLP Tasks (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Recent studies have focused on rule-based and neural sequence-to-sequence (seq2sequ) TS is a technique that reduces text complexity for human consumption. |
| Approach: | They evaluate two possible uses of neural TS: simplifying input texts at prediction time and augmenting training data to provide machines with additional information during training. |
| Outcome: | The proposed approach improves performance on two datasets. |
An Unsupervised Method for Learning Representations of Multi-word Expressions for Semantic Classification (2020.coling-main)
Copied to clipboard
| Challenge: | Existing methods for learning multi-word expressions have language sparsity and are not supervised. |
| Approach: | They propose an unsupervised approach to learning a compositional representation function for multi-word expressions . they use a Tratz dataset to train the composition function on the word-semantic relation . |
| Outcome: | The proposed method outperforms the previous state-of-the-art method on the Tratz dataset with an F1 score of 50.4%. |
Bootstrapping Polar-Opposite Emotion Dimensions from Online Reviews (L18-1)
Copied to clipboard
| Challenge: | Existing bootstrapping methods for learning lexicons from unannotated online texts have important drawbacks. |
| Approach: | They propose a bootstrapping approach that softly labels unlabeled terms for polar-opposite emotion dimension values from the Ortony/Clore/Collins model of emotions. |
| Outcome: | The proposed approach achieves considerably better performance than several baselines. |
Can LLMs Judge Debates? Evaluating Non-Linear Reasoning via Argumentation Theory Semantics (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large Language Models excel at linear reasoning tasks but are underexplored on non-linear structures such as natural debates. |
| Approach: | They evaluate whether Large Language Models can approximate structured reasoning from Computational Argumentation Theory. |
| Outcome: | The proposed model performs well on dialogue-formatted debates without access to the underlying graph. |
Alignment over Heterogeneous Embeddings for Question Answering (N19-1)
Copied to clipboard
| Challenge: | Existing approaches for non-factoid question answering are based on heterogeneous embeddings that model text at different levels of abstraction. |
| Approach: | They propose a fast, mostly-unsupervised approach for non-factoid question answering called Alignment over Heterogeneous Embeddings (AHE) it aligns each word in the question and candidate answer with the most similar word in retrieved supporting paragraph and a meta-classifier that learns how much to trust the predictions over each representation. |
| Outcome: | The proposed approach outperforms other supervised approaches on the AI2 Reasoning Challenge dataset and the WikiQA dataset. |
Students Who Study Together Learn Better: On the Importance of Collective Knowledge Distillation for Domain Transfer in Fact Verification (2021.emnlp-main)
Copied to clipboard
| Challenge: | Neural networks depend heavily on lexicalized information, which can be overfitted . this can be a problem in fact verification, which has important societal implications. |
| Approach: | They propose a knowledge distillation approach for fact verification using student models. |
| Outcome: | The proposed approach outperforms state-of-the-art classifiers on a training dataset and in supervised settings. |
Hiding in Plain Sight: Tweets with Hate Speech Masked by Homoglyphs (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Currently, there are almost 150,000 Unicode characters, which presents extensive substitution possibilities. |
| Approach: | They develop a character substitution scraping method to collect hate speech . they use an annotated dataset with 1,281 non-Latin characters to scrape out offensive words . |
| Outcome: | The proposed method can detect hate speech with annotated data, but it performs poorly in a zero-shot setting. |
A STEP towards Interpretable Multi-Hop Reasoning:Bridge Phrase Identification and Query Expansion (2022.lrec-1)
Copied to clipboard
| Challenge: | Identifying bridge phrases remains one of the challenges for multi-hop question answering . |
| Approach: | They propose an unsupervised method for the identification of bridge phrases in multi-hop question answering . they construct a graph of noun phrases from the question and available context . |
| Outcome: | The proposed method improves all downstream components in a multi-hop QA system. |