• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • 3
  • Tagged with
  • 3
  • 3
  • 2
  • 2
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • About
  • The Global ETD Search service is a free service for researchers to find electronic theses and dissertations. This service is provided by the Networked Digital Library of Theses and Dissertations.
    Our metadata is collected from universities around the world. If you manage a university/consortium/country archive and want to be added, details can be found on the NDLTD website.
1

Parameter-efficient modeling and robust automatic evaluation of image captioning

Ahmadi, Saba 10 1900 (has links)
Le sous-titrage d’images est la tâche de l’intelligence artificielle (IA) qui consiste à décrire des images en langage naturel. Cette tâche d’IA a plusieurs applications sociétales utiles, telles que l’accessibilité pour les malvoyants, la génération automatisée de contenu, l’interaction humain-robot et l’analyse d’imagerie médicale. Au cours des huit dernières années, la recherche sur le sous-titrage d'images a connu d'énormes progrès dans la création de modèles solides, la collecte d'ensembles de données à grande échelle ainsi que le développement de mesures d'évaluation automatique. Malgré ces progrès remarquables, la recherche sur le sous-titrage d'images est confrontée à deux défis majeurs: 1) Comment construire des modèles efficaces en termes de paramètres, et 2) Comment construire des métriques d'évaluation automatique robustes. Dans cette thèse, nous apportons notre contribution à la résolution de chacun de ces défis. Premièrement, nous proposons une méthode efficace en termes de paramètres (MAPL \cite{mapl}) qui adapte des modèles pré-entraînés unimodaux de vision uniquement et de langage uniquement pour la tâche multimodale de sous-titrage d'images. MAPL apprend un mappage léger entre les espaces de représentation des modèles unimodaux. Ainsi, MAPL peut exploiter les fortes capacités de généralisation des modèles unimodaux pré-entraînés pour des tâches multimodales telles que le sous-titrage d'images. Deuxièmement, nous présentons une étude systématique de la robustesse des mesures d’évaluation des sous-titres d’images récemment proposées. Même si ces métriques correspondent bien aux jugements humains, nous avons constaté qu'elles ne sont pas robustes pour identifier les erreurs fines dans les légendes générées par le modèle. Il faut donc faire preuve de prudence lors de l'utilisation de ces métriques pour l'évaluation des sous-titres d'images. Nous espérons que nos résultats guideront de nouvelles améliorations dans l’évaluation automatique du sous-titrage d’images. / Image captioning is the artificial intelligence (AI) task of describing images in natural language. This AI task has several useful societal applications, such as accessibility for the visually impaired, automated content generation, human-robot interaction, and medical imaging analysis. Over the last eight years, image captioning research has seen tremendous progress in building strong models, collecting large scale datasets as well as developing automatic evaluation metrics. Despite such remarkable progress, image captioning research faces two major challenges: 1) How to build parameter-efficient models, and 2) How to build robust automatic evaluation metrics. In this thesis, we make contributions towards tackling each of these challenges. First, we propose a parameter efficient method (MAPL \cite{mapl}) that adapts pre-trained unimodal vision-only and language-only models for the multimodal task of image captioning. MAPL learns a lightweight mapping between the representation spaces of the unimodal models. Thus, MAPL can leverage the strong generalization capabilities of the pre-trained unimodal models for multimodal tasks such as image captioning. Second, we present a systematic study of the robustness of recently proposed image captioning evaluation metrics. Even though these metrics correlate well with human judgments, we found that these metrics are not robust in identifying fine-grained errors in model generated captions, and thus, caution needs to be exercised when using these metrics for image captioning evaluation. We hope our findings will guide further improvements in the automatic evaluation of image captioning.
2

Parameter efficiency in Fine tuning Pretrained Large Language Models for Downstream Tasks

Dorairaj, Jonathan January 2024 (has links)
This thesis investigates Parameter-Efficient Fine-Tuning (PEFT) methods, specifically Low-Rank Adaptation (LoRA) (Hu et al. 2021) and Adapters (Houlsby et al. 2019), using the General Language Understanding Evaluation (GLUE) dataset (Wang et al. 2019). The primary focus is to evaluate the effectiveness and efficiency of these methods in fine-tuning pre-trained language models. Additionally, we introduce a novel application by applying the methodology from Yang et al. 2024 to the adapter module weights. We utilize Laplace approximations over both the LoRA (Yang et al. 2024, Daxberger et al. 2022a) and the newly adapted Adapter weights, assessing the Expected Calibration Error (ECE) and Negative Log-Likelihood (NLL). Furthermore, we discuss practical considerations such as training time, memory usage, and storage space implications of these PEFT techniques. The findings provide valuable insights into the trade-offs and benefits of using LoRA and Adapters for fine-tuning in resource-constrained environments.
3

Automatic text summarization of French judicial data with pre-trained language models, evaluated by content and factuality metrics

Adler, Malo January 2024 (has links)
During an investigation carried out by a police officer or a gendarme, audition reports are written, the length of which can be up to several pages. The high-level goal of this thesis is to study various automatic and reliable text summarization methods to help with this time-consuming task. One challenge comes from the specific, French and judicial data that we wish to summarize; and another challenge comes from the need for reliable and factual models. First, this thesis focuses on automatic summarization evaluation, in terms of both content (how well the summary captures essential information of the source text) and factuality (to what extent the summary only includes information from or coherent with the source text). Factuality evaluation, in particular, is of crucial interest when using LLMs for judicial purposes, because of their hallucination risks. Notably, we propose a light variation of SelfCheckGPT, which has a stronger correlation with human judgment (0.743) than the wide-spread BARTScore (0.542), or our study dataset. Other paradigms, such as Question-Answering, are studied in this thesis, which however underperform compared to these. Then, extractive summarization methods are explored and compared, including one based on graphs via the TextRank algorithm, and one based on greedy optimization. The latter (overlap rate: 0.190, semantic similarity: 0.513) clearly outperforms the base TextRank (overlap rate: 0.172, semantic similarity: 0.506). An improvement of the TextRank with a threshold mechanism is also proposed, leading to a non-negligible improvement (overlap rate: 0.180, semantic similarity: 0.513). Finally, abstractive summarization, with pre-trained LLMs based on a Transformer architecture, is studied. In particular, several general-purpose and multilingual models (Llama-2, Mistral and Mixtral) were objectively compared on a summarization dataset of judicial procedures from the French police. Results show that the performances of these models are highly related to their size: Llama-2 7B struggles to adapt to uncommon data (overlap rate: 0.083, BARTScore: -3.099), while Llama-2 13B (overlap rate: 0.159, BARTScore: -2.718) and Llama-2 70B (overlap rate: 0.191, BARTScore: -2.479) have proven quite versatile and efficient. To improve the performances of the smallest models, empirical prompt-engineering and parameter-efficient fine-tuning are explored. Notably, our fine-tuned version of Mistral 7B reaches performances comparable to those of much larger models (overlap rate: 0.185, BARTScore: -2.060), without the need for empirical prompt-engineering, and with a linguistic style closer to what is expected. / Under en utredning som görs av en polis eller en gendarm skrivs förhörsprotokoll vars längd kan vara upp till flera sidor. Målet på hög nivå med denna rapport är att studera olika automatiska och tillförlitliga textsammanfattningsmetoder för att hjälpa till med denna tidskrävande uppgift. En utmaning kommer från de specifika franska och rättsliga uppgifter som vi vill sammanfatta; och en annan utmaning kommer från behovet av pålitliga, sakliga och uppfinningsfria modeller. För det första fokuserar denna rapport på automatisk sammanfattningsutvärdering, både vad gäller innehåll (hur väl sammanfattningen fångar väsentlig information i källtexten) och fakta (i vilken utsträckning sammanfattningen endast innehåller information från eller överensstämmer med källtexten). Faktautvärdering, i synnerhet, är av avgörande intresse när man använder LLM för rättsliga ändamål, på grund av deras hallucinationsrisker. Vi föreslår särskilt en lätt variant av SelfCheckGPT, som har en starkare korrelation med mänskligt omdöme (0,743) än den utbredda BARTScore (0,542), eller vår studiedatauppsättning. Andra paradigm, såsom Question-Answering, studeras i denna rapport, som dock underpresterar jämfört med dessa. Sedan utforskas och jämförs extraktiva sammanfattningsmetoder, inklusive en baserad på grafer via TextRank-algoritmen och en baserad på girig optimering. Den senare (överlappning: 0,190, semantisk likhet: 0,513) överträffar klart basen TextRank (överlappning: 0,172, semantisk likhet: 0,506). En förbättring av TextRank med en tröskelmekanism föreslås också, vilket leder till en icke försumbar förbättring (överlappning: 0,180, semantisk likhet: 0,513). Slutligen studeras abstrakt sammanfattning, med förutbildade LLM baserade på en transformatorarkitektur. I synnerhet jämfördes flera allmänna och flerspråkiga modeller (Llama-2, Mistral och Mixtral) objektivt på en sammanfattningsdatauppsättning av rättsliga förfaranden från den franska polisen. Resultaten visar att prestandan för dessa modeller är starkt relaterade till deras storlek: Llama-2 7B kämpar för att anpassa sig till ovanliga data (överlappning: 0,083, BARTScore: -3,099), medan Llama-2 13B (överlappning: 0,159, BARTScore: -2,718) och Llama-2 70B (överlappning: 0,191, BARTScore: -2,479) har visat sig vara ganska mångsidiga och effektiva. För att förbättra prestandan för de minsta modellerna utforskas empirisk prompt-teknik och parametereffektiv finjustering. Noterbart är att vår finjusterade version av Mistral 7B når prestanda som är jämförbara med de för mycket större modeller (överlappning: 0,185, BARTScore: -2,060), utan behov av empirisk prompt-teknik och med en språklig stil som ligger närmare vad som förväntas.

Page generated in 0.0686 seconds