Large Language Models (LLMs) have demonstrated impressive capability to solve a wide range of tasks in recent years. This has inspired researchers and practitioners in the real-world industrial domain to build useful products via leveraging LLMs. However, extensive evaluations of LLMs, in terms of accuracy, memory management, and inference latency, while ensuring the reproducibility of the results are crucial before deploying LLM-based solutions for real-world usage. In addition, when evaluating LLMs on internal customer data, an on-premise evaluation system is necessary to protect customer privacy rather than sending customer data to third-party APIs for evaluation. In this paper, we demonstrate how we build an on-premise system for LLM evaluation to address the challenges in the evaluation of LLMs in real-world industrial settings. We demonstrate the complexities of consolidating various datasets, models, and inference-related artifacts in complex LLM inference pipelines. For this purpose, we also present a case study in a real-world industrial setting. The demonstration of the LLM evaluation tool development would help researchers and practitioners in building on-premise systems for LLM evaluation ensuring privacy, reliability, robustness, and reproducibility.

Nous présentons un modèle d’apprentissage automatique qui combine modèles neuronaux et linguistiques pour traiter les tâches de classification dans lesquelles la distribution des étiquettes des instances est déséquilibrée. Les performances de ce modèle sont mesurées à l’aide d’expériences menées sur les tâches de classification de recettes de cuisine de la campagne DEFT 2013 (Grouin et al., 2013). Nous montrons que les plongements lexicaux (word embeddings) associés à des méthodes d’apprentissage profond obtiennent de meilleures performances que tous les algorithmes déployés lors de la campagne DEFT. Nous montrons aussi que ces mêmes classifieurs avec plongements lexicaux peuvent gagner en performance lorsqu’un modèle linguistique est ajouté au modèle neuronal. Nous observons que l’ajout d’un modèle linguistique au modèle neuronal améliore les performances de classification sur les classes rares.

In this paper, we report on our experiments towards the creation of a corpus for coherence evaluation. Most corpora for textual coherence evaluation are composed of randomly shuffled sentences that focus on sentence ordering, regardless of whether the sentences were originally related by a discourse relation. To the best of our knowledge, no publicly available corpus has been designed specifically for the evaluation of coherence of known discursive units. In this paper, we focus on coherence modeling at the intra-discursive level and describe our approach to build a corpus of incoherent pairs of sentences. We experimented with a variety of corruption strategies to create synthetic incoherent pairs of discourse arguments from coherent ones. Using discourse argument pairs from the Penn Discourse Tree Bank, we generate incoherent discourse argument pairs, by swapping either their discourse connective or a discourse argument. To evaluate how incoherent the generated corpora are, we use a convolutional neural network to try to distinguish the original pairs from the corrupted ones. Results of the classifier as well as a manual inspection of the corpora show that generating such corpora is still a challenge as the generated instances are clearly not “incoherent enough”, indicating that more effort should be spent on developing more robust ways of generating incoherent corpora.

In this paper, we propose a neural-based model to address the first task of the DEFT 2013 shared task, with the main challenge of a highly imbalanced dataset, using state-of-the-art embedding approaches and deep architectures. We report on our experiments on the use of linguistic features, extracted by Charton et. al. (2014), in different neural models utilizing pretrained embeddings. Our results show that all of the models that use linguistic features outperform their counterpart models that only use pretrained embeddings. The best performing model uses pretrained CamemBERT embeddings as input and CNN as the hidden layer, and uses additional linguistic features. Adding the linguistic features to this model improves its performance by 4.5% and 11.4% in terms of micro and macro F1 scores, respectively, leading to state-of-the-art results and an improved classification of the rare classes.

This paper describes a new approach for the task of contextual emotion detection. The approach is based on a neural feature extractor, composed of a recurrent neural network with an attention mechanism, followed by a classifier, that can be neural or SVM-based. We evaluated the model with the dataset of the task 3 of SemEval 2019 (EmoContext), which includes short 3-turn conversations, tagged with 4 emotion classes. The best performing setup was achieved using ELMo word embeddings and POS tags as input, bidirectional GRU as hidden units, and an SVM as the final classifier. This configuration reached 69.93% in terms of micro-average F1 score on the main 3 emotion classes, a score that outperformed the baseline system by 11.25%.

This paper describes our system at SemEval 2019, Task 3 (EmoContext), which focused on the contextual detection of emotions in a dataset of 3-round dialogues. For our final system, we used a neural network with pretrained ELMo word embeddings and POS tags as input, GRUs as hidden units, an attention mechanism to capture representations of the dialogues, and an SVM classifier which used the learned network representations to perform the task of multi-class classification. This system yielded a micro-averaged F1 score of 0.7072 for the three emotion classes, improving the baseline by approximately 12%.

This paper summarizes our participation to the CLPsych 2019 shared task, under the name CLaC. The goal of the shared task was to detect and assess suicide risk based on a collection of online posts. For our participation, we used an ensemble method which utilizes 8 neural sub-models to extract neural features and predict class probabilities, which are then used by an SVM classifier. Our team ranked first in 2 out of the 3 tasks (tasks A and C).

Native language identification (NLI) is the task of determining an author’s native language, based on a piece of his/her writing in a second language. In recent years, NLI has received much attention due to its challenging nature and its applications in language pedagogy and forensic linguistics. We participated in the NLI2017 shared task under the name UT-DSP. In our effort to implement a method for native language identification, we made use of a fusion of character and word N-grams, and achieved an optimal F1-Score of 77.64%, using both essay and speech transcription datasets.