NeoAMT: Neologism-Aware Agentic Machine Translation with Reinforcement Learning

TL;DR

NeoAMT addresses neologism translation by coupling an agentic translation agent with a Wiktionary-based dictionary retrieval tool. It introduces Neko, a multilingual benchmark (16 languages, 75 directions) derived from Wiktionary, and a retrieval-enabled RL framework with a novel reward design and translation-difficulty-aware adaptive rollout. Empirical results show limited gains from SFT, advantages of RL with reasoning and dictionary search over baselines, and nuanced insights from human evaluation and thinking-path analyses. The work demonstrates that integrating external lexical resources into reinforcement learning can improve neologism handling, while also highlighting challenges in retrieval quality and prompt fidelity that shape future improvements.

Abstract

Neologism-aware machine translation aims to translate source sentences containing neologisms into target languages. This field remains underexplored compared with general machine translation (MT). In this paper, we propose an agentic framework, NeoAMT, for neologism-aware machine translation using a Wiktionary search tool. Specifically, we first create a new dataset for neologism-aware machine translation and develop a search tool based on Wiktionary. The new dataset covers 16 languages and 75 translation directions and is derived from approximately 10 million records of an English Wiktionary dump. The retrieval corpus of the search tool is also constructed from around 3 million cleaned records of the Wiktionary dump. We then use it for training the translation agent with reinforcement learning (RL) and evaluating the accuracy of neologism-aware machine translation. Based on this, we also propose an RL training framework that contains a novel reward design and an adaptive rollout generation approach by leveraging "translation difficulty" to further improve the translation quality of translation agents using our search tool.

Figures (7)

• Figure 1: A Chinese-English neologism-aware machine translation example from our Neko dataset. The text in blue highlights a neologism or its translation. The source Chinese text contains a neologism "给她爱" which means "GTA", an action-adventure video game series. The circle denotes that the thinking and searching processes can be repeated. The detailed thinking and searching process of our model can be found in Table \ref{['tab:find_new_meaning_example_our_model_gta']}, Appendix \ref{['app:ex']}.
• Figure 2: The construction process of our Neko benchmark. We clean about 10M records from a Wiktionary dump. In each cleaned word entry, we have word, part-of-speech, etymology, senses and non-disambiguated translations. In word senses, we have tags, such as "neologism", "Internet", glosses and so on. Glosses contain definitions of words. Example sentences and translations are available for some words. We obtain three types of Wiktionary word entries from the cleaned records. Blue words are neologisms or their translations.
• Figure 3: Statistics of the test split of Neko dataset and the constructed dictionary in the search toolkit.
• Figure 4: Screenshot of translation ranking annotation application.
• Figure 5: The comparison of XCOMET scores on the WMT24pp test dataset grouped by language pairs between before and after our training using the Neko dataset.

Theorems & Definitions (1)

• Definition 1: Translation Difficulty