The Thesis in Twenty Terms

As a small application of the lexical retrieval methods discussed in this thesis, this chapter gives a compact view of the vocabulary that most distinguishes each main-body chapter. The lists below contain the twenty terms with the highest term frequency–inverse document frequency (TF–IDF) weight in each chapter ¹Spärck Jones (1972)A Statistical Interpretation of Term Specificity and Its Application in Retrieval. TF–IDF and other recurring technical terms are defined in Appendix.

The computation treats each of the eight main-body chapters as one document. The text is normalized by removing comments, LaTeX commands, citations, references, mathematics, English stopwords, and common document scaffolding. Words are converted to lowercase unaccented forms, and unigram terms of at least three characters are retained. These exclusions are deliberately limited to generic structural or formatting artifacts rather than topic-specific words. For a term $t$ in chapter $d$, the weight is $$\operatorname{tfidf}(t,d) = \frac{f_{t,d}}{\sum_{t'} f_{t',d}} \left( \log\frac{1 + N}{1 + n_t} + 1 \right),$$ where $N=8$ is the number of chapters and $n_t$ is the number of chapters containing $t$. The smoothing avoids zero weights, and alphabetical order breaks exact score ties.

Introduction

retrieval, information, document, visual, search, thesis, models, croissantllm, author, systems, documents, model, work, series, colpali, knowledge, language, performance, efforts, vidore.

A Brief History of Information Retrieval

retrieval, document, documents, ranking, information, query, systems, term, probabilistic, models, search, sparse, embedding, lexical, modern, became, early, one, evidence, vectors.

ColPali

retrieval, document, visual, image, models, paligemma, colpali, performance, patches, model, embeddings, query, text, documents, unstructured, systems, tasks, vidore, vision, training.

ViDoRe Benchmarks: V2 & V3

visual, queries, query, retrieval, grounding, bounding, vidore, annotators, annotation, human, textual, generation, answer, document, boxes, pro, summaries, pages, gemini, colembed.

ModernVBERT

image, retrieval, models, visual, document, training, modality, encoder, model, alignment, performance, tasks, contrastive, bidirectional, vision, attention, causal, text, resolution, tokens.

Contextual Document Embeddings

chunks, chunking, chunk, retrieval, contextual, document, conteb, method, training, information, models, embedding, context, documents, performance, late, insent, approaches, modernbert, queries.

CroissantLLM

model, data, french, croissantllm, models, english, training, performance, tokens, language, llama, code, bloom, dataset, transparency, corpus, license, translation, trained, shot.

Personal Perspectives on Search

memory, search, retrieval, information, document, model, associative, parametric, working, evidence, systems, stores, external, links, operation, market, llm, attention, index, agentic.