flâneur

Avery Chen

3 followers · 3 following · 289 views

on the atlas — 30

highlights — 43

  • to evaluate the copy rate of systems, i.e. their ab
    2404.02835
  • In high-density domains, it is thus easier to retrieve relevant translation examples
    2404.02835
  • unpro- ductive to retrieve multiple identical examples
    2404.02835
  • We notably find that (a) retrieval actually matters for edit-based and in-context learning; (b) existing retrieval pipelines can be simplified at inference; (c) optimizing source coverage and/or instance diver- sity is helping, especially when the closest match is poor
    2404.02835
  • understand the interplay between the retrieval and generation tasks,
    2404.02835
  • the effects of increasing the number and di- versity of examples,
    2404.02835
  • varying re- trieval methods for several translation archi- tectures,
    2404.02835
  • the up- stream retrieval step is mostly unexplored
    2404.02835
  • retrieve the top- k most relevant en- tities E t = { e 1 , . . . , e k } from a knowledge graph
    2024.emnlp-main.914.pdf
  • leverage an external knowledge source to first retrieve the most relevant entities for an input text, and then gener- ate the translation by incorporating the retrieved entity names in the target language
    2024.emnlp-main.914.pdf
  • it does not necessarily account for the transcre- ation of entity names between languages using the same script, like English and Italian
    2024.emnlp-main.914.pdf
  • adapting an entity name from the script of one language to another
    2024.emnlp-main.914.pdf
  • multilingual MT systems that can translate to and from multiple languages with a single model,
    2024.emnlp-main.914.pdf
  • hen and how to translate entities whose names are significantly different across lan- guages
    2024.emnlp-main.914.pdf
  • Indeed, word-for-word , or literal , translations of their names may not be suitable due to cultural-specific references, which can vary de- pending on social, geographical, historical, and po- litical contexts, among other factors
    2024.emnlp-main.914.pdf
  • hese de- velopments have led to MT systems that not only perform exceptionally well in high-resource lan- guages but also support a growing number of low- resource language
    2024.emnlp-main.914.pdf
  • Translating text that contains entity names is a challenging task, as cultural-related references can vary significantly across languages. These variations may also be caused by transcre- ation , an adaptation process that entails more than transliteration and word-for-word trans- lation.
    2024.emnlp-main.914.pdf
  • To achieve better translation, we let GPT-4o first judge the relevance of the given document to respond with a judgment and a 5-point rating, and then translate the sentence in a chain-of-thought (CoT) manner.
    2025.findings-emnlp.313.pdf
  • Given this, the consecutive paragraphs from D l \ p l 1 ( l ∈{ zh , de , fr , cs } ) form as the relevant document in other languages
    2025.findings-emnlp.313.pdf
  • involve knowledge- intensive semantics
    2025.findings-emnlp.313.pdf
  • The multi-task training samples are low-cost, and do not require additional labeling costs
    2025.findings-emnlp.313.pdf
  • cross-lingual information completion, self-knowledge-enhanced translation and cross-lingual relevance discrimina- tion.
    2025.findings-emnlp.313.pdf
  • the model might receive multiple documents from various lan- guages. These multilingual documents are not re- stricted to parallel documents and can convey di- verse meanings
    2025.findings-emnlp.313.pdf
  • we first build a benchmark dataset, named RAGtrans .
    2025.findings-emnlp.313.pdf
  • (2) Retrieving knowledge triplets : the others retrieve relevant information from knowledge graphs to let the models know domain or cultural knowledge w.r.t. the source sentences
    2025.findings-emnlp.313.pdf
  • (1) Retrieving in-context examples (also known as “translation memory”): for a source sentence, a few studies retrieve the relevant paired sentences from bilingual corpora to enhance MT models
    2025.findings-emnlp.313.pdf
  • With the help of retrieved information, LLMs could generate more accurate and knowledge-enrich responses (Li et al., 2022; Gao et al., 2023)
    2025.findings-emnlp.313.pdf
  • training method to teach LLMs how to use information from multilingual docu- ments during their translation.
    2025.findings-emnlp.313.pdf
  • unstructured documents, and might not be fully paired across different languages
    2025.findings-emnlp.313.pdf
  • etrieves in-context examples from paired MT corpora, or domain-specific knowledge from knowledge graphs, to enhance MT models.
    2025.findings-emnlp.313.pdf
  • Chain-of-Thought (CoT) and Tree- of-Thought (ToT) prompting incur high computa- tional costs due to increased token usage (Figure 4), resulting in substantial API expenses (Figure 9). This may hinder accessibility, particularly for re- searchers with limited resources
    2025.acl-srw.17.pdf
  • parallel * Equal contribution. † Equal advising. 1 We make our code and data publicly available at: https: //github.com/cisnlp/manchu-in-context-mt . data is scarce or even unavailable for many low- resource or endangered languages (Haddow et al., 2022), which prevents the training of dedicated MT systems for these languages
    2025.acl-long.429.pdf
  • generating synthetic parallel data through in-context MT offers a pathway to mit- igate data scarcity and build effective and effi- cient low-resource neural MT systems
    2025.acl-long.429.pdf
  • how each resource and its qual- ity affect the translation performance,
    2025.acl-long.429.pdf
  • read- ily take advantage of linguistic resources such as grammar books and dictionaries.
    2025.acl-long.429.pdf
  • his research establishes RARE as a scalable solution for improving LLMs in domains where logical coherence and factual integrity are critica
    2025.acl-long.896.pdf
  • for complex, knowledge- intensive tasks such as medical and common- sense reasoning
    2025.acl-long.896.pdf
  • Tree-of-Thought (ToT) prompting (Yao et al., 2023) extends this concept by enabling exploration of multiple translation candidates simultaneously
    2025.acl-srw.17.pdf
  • Chain-of- Thought (CoT) prompting (Wei et al., 2022b) breaks down complex reasoning into intermedi- ate steps
    2025.acl-srw.17.pdf
  • The effectiveness of few- shot prompting depends critically on three factors: (1) the quality of provided examples, (2) their di- versity across linguistic constructions, and (3) their relevance to the target domain.
    2025.acl-srw.17.pdf
  • bility without the computational cost of retraining. However, challenges remain in optimizing these prompting approaches (Yao et al., 2023; Zhang et al., 2024), ensuring consistent quality across di- verse language combinations (Zhu et al., 2024; Xie et al., 2023), and addressing the computational de- mands of inference with large models (Xia et al., 2024; Bapna and Firat, 2019). 2.2 Prompting Strategies for Translation Prompting strategies fundamentally shape how LLMs approach translation tasks, offering different trade-offs between simplicity, performance, and computational efficiency. We ex…
    2025.acl-srw.17.pdf
  • Zero-shot prompting leverages an LLM’s pre- trained knowledge to perform translations without any task-specific examples (Brown et al., 2020). This approach relies entirely on the model’s exist- ing parameters, making its effectiveness heavily dependent on the language pair’s representation in the pre-training corpus (Vilar et al., 2023). While effective for high-resource languages, zero-shot translation often falters with idiomatic expressions, rare vocabulary, and specialized terminology (Jiao et al., 2023)
    2025.acl-srw.17.pdf
  • LLMs excel at incorporating contextual information and maintaining semantic consistency across languages (Zhu et al., 2024; Gar- cia et al., 2023), though their performance varies substantially across language pairs (Sanh et al., 2022; Zhang et al., 2023). High-resource languages typically benefit from better representation in pre- training data (Kudugunta et al., 2023; Costa-jussà et al., 2022), while low-resource languages often present ongoing challenges
    2025.acl-srw.17.pdf