Named Entity Recognition (NER) in the tender and procurement domain is critical for tasks such as contract monitoring, supplier analysis, and compliance tracking. However, unlike general-purpose NER, no open-source datasets exist for Tender NER, largely due to data sensitivity and confidentiality restrictions. This scarcity limits the development of automated entity extraction models. To address this gap, we propose struct2unstruct, a data preparation pipeline that generates and annotates tender-specific datasets using large language models (LLMs). Starting from structured procurement data published by the Singapore government (2015–2021) available in English language, we employ Llama-3 to generate synthetic tender narratives in multiple writing styles, ensuring each contains at least one tender-related entity. Post-processing steps correct inconsistencies in dates, symbols, and entity formats. Entities are then annotated using a BIO tagging scheme through deterministic alignment with structured fields, followed by expert validation to ensure accuracy. This study focuses on data preparation and evaluation, not model training. The resulting dataset provides a scalable resource for future Tender NER research in low-resource environments. By releasing both the dataset and pipeline as open-source resources, we establish a foundation for advancing domain-adapted information extraction and automated tender entity recognition.

Speech emotion recognition (SER) has gained significant interest in recent times, which utilizes speech signals to identify the emotional state of speakers. Accurate recognition of subtle emotional variations in speech, such as distinguishing closely related emotional states, remains a challenging problem due to the variability of speech signals and the acoustic similarity among emotion classes across different speakers and linguistic contexts. This paper proposes a hybrid deep learning model that integrates a Convolutional Neural Network (CNN) with a Bidirectional Long Short-Term Memory (BiLSTM) network to effectively identify both spectral and temporal features of speech. The log Mel-frequency spectral coefficients (MFSC) are used as input features to represent discriminative spectral representations, while the BiLSTM layer model represents long-range temporal dependencies in speech signals. The proposed framework is evaluated on the Toronto Emotional Speech Set (TESS), a publicly available dataset of acted emotional speech containing seven emotion classes. The experimental findings show that the hybrid CNN-BiLSTM achieved an overall classification accuracy of 96.36%, significantly outperforming baseline models including GRU (91.84%), BiLSTM (93.12%), and CNN–GRU (94.67%). These findings highlight the effectiveness of combining spectral and temporal modeling for improved speech emotion recognition performance. Furthermore, our CNN+BiLSTM approach offers a computationally efficient and data-efficient alternative to transformer-based models, while still effectively capturing both spatial and temporal emotional cues in speech, making it suitable for real-time and resource-constrained applications.

Extracting structured text from complex tables in PDF tender documents remains a challenging task due to the loss of structural and positional information during the extraction process. AI-based models often require extensive training data, making development from scratch both tedious and time-consuming. Our research focuses on identifying tender entities in complex table formats within PDF documents. To address this, we propose a novel approach utilizing few-shot learning with large language models (LLMs) to restore the structure of extracted text. Additionally, handcrafted rules and regular expressions are employed for precise entity classification. To evaluate the robustness of LLMs with few-shot learning, we employ data-shuffling techniques. Our experiments show that current text extraction tools fail to deliver satisfactory results for complex table structures. However, the few-shot learning approach significantly enhances the structural integrity of extracted data and improves the accuracy of tender entity identification.

Multiword expressions, characterised by non-compositional meanings and syntactic irregularities, are an example of nuanced language. These expressions can be used literally or idiomatically, leading to significant changes in meaning. Although large language models perform well on many tasks, their ability to handle subtle linguistic phenomena remains unclear. This study examines how state-of-the-art models process the ambiguity of potentially idiomatic multiword expressions, particularly in less frequent contexts where memorisation is less likely to help. By evaluating models in Portuguese, Galician, and English, and introducing a new code-switched dataset and task, we show that large language models, despite their strengths, have difficulty handling nuanced language. In particular, we find that the latest models, including GPT-4, fail to outperform the xlm-roBERTa-base baselines in both detection and semantic tasks, with especially poor performance on the novel tasks we introduce, despite its similarity to existing tasks. Overall, our results demonstrate that multiword expressions, especially those that are ambiguous, continue to be a challenge to models. We provide open access to our datasets, prompts and model responses.

In the public procurement domain, extracting accurate tender entities from unstructured text remains a critical, less explored challenge, because tender data is highly sensitive and confidential, and not available openly. Previously, state-of-the-art NLP models were developed for this task; however developing an NER model from scratch required huge amounts of data and resources. Similarly, performing fine-tuning of a transformer-based model like BERT requires training data, as a result posing challenges in training data cost, model generalization, and data privacy. To address these challenges, an emerging LLM such as GPT-4 in a Few-shot learning environment achieves SOTA performance comparable to fine-tuned models. However, being dependent on the closed-source commercial LLMs involves high cost and privacy concerns. In this study, we have investigated open-source LLMs like Mistral and LLAMA-3, focusing on the tender domain for the NER tasks on local consumer-grade CPUs in three different environments: Zero-shot, One-shot, and Few-shot learning. The motivation is to efficiently lessen costs compared to a cloud solution while preserving accuracy and data privacy. Similarly, we have utilized two datasets open-source from Singapore and closed-source commercially sensitive data provided by Siemens. As a result, all the open-source LLMs achieve above 85% F1-score on an open-source dataset and above 90% F1-score on a closed-source dataset.

In this paper, we present the system proposed by our team OldJoe, for the 8th edition of the AVeriTeC shared task, as part of the FEVER workshop. The objective of this task is to verify the factuality of real-world claims. Our approach integrates open source large language models, SQL, and in-context learning. We begin with embedding the knowledge store using a pretrained embedding language model then storing the outputs in a SQL database. Subsequently, we prompt an LLM to craft relevant questions based on the input claim, which are then used to guide the retrieval process. We further prompt the LLM to generate answers to the questions and predict the veracity of the original claim. Our system scored 0.49 on the HU-METEOR AVeriTeC score on the dev set and 0.15 on the Ev2R recall on the test set. Due to the time constraint we were unable to conduct additional experiments or further hyperparameter tuning. As a result, we adopted this pipeline configuration centered on the Qwen3-14B-AWQ model as our final submission strategy. The full pipeline is available on GitHub: https://github.com/farahft/OldJoe