TL;DR
This paper introduces Wiki-MT, a system that leverages multilingual knowledge graphs like Wikidata to improve machine translation (MT) of entity names. By retrieving human-curated aliases and translations from Wikidata, Wiki-MT produces more accurate translations and offers users insights into the translation process, enhancing both quality and interpretability.
What Problem Does This Paper Address?
Machine translation systems have made incredible strides, but they still falter when translating entity names, especially those with cultural nuances or non-literal translations (transcreation). For instance, translating a movie title like "The Catcher in the Rye" into Italian requires more than a word-for-word approach; it becomes "Il giovane Holden." Current MT systems often try to address this by augmenting training data with vast numbers of entity name translations. However, this approach is inefficient, leading to unmanageably large datasets and requiring constant retraining as new entities emerge or existing ones are updated. This paper tackles the challenge of making entity name translation more robust, scalable, and interpretable.
Methods
Wiki-MT is built upon two core components:
- Knowledge Retriever: This component takes an input query and uses vector representations to find the top-3 most relevant entities from Wikidata. It achieves this by encoding the entity's name and description to create a vector representation and then calculating cosine similarity with the query's vector representation.
- Knowledge-Enhanced Machine Translator: This is a standard MT system (specifically, NLLB-200) that is augmented with retrieved knowledge. For each relevant entity identified by the retriever, its name in the source language and its translated name in the target language are appended to the original query using special tokens (
[META]and[AS]). This augmented query is then fed into the MT system, which has been fine-tuned on a dataset of knowledge-seeking queries (Mintaka) to better utilize this contextual information.
Datasets
- Wikidata: Used as the source of multilingual knowledge for entity names, descriptions, and translations. The paper mentions fine-tuning the knowledge retriever on a Wikidata dump from May 2023 covering 11 languages.
- Mintaka: A dataset of knowledge-seeking queries used for fine-tuning the knowledge-enhanced MT system.
Key Insights and Ideas
- Knowledge Graphs as a Scalable Alternative to Data Augmentation: Instead of relying on massive, static training datasets for entity name translation, integrating knowledge graphs offers a dynamic and more manageable solution. Updating a knowledge graph is significantly easier and faster than retraining an entire MT model.
- Bridging the Gap Between Literal Translation and Transcreation: By providing human-curated aliases and translations from knowledge graphs, Wiki-MT can better handle cases where a literal translation would be incorrect or misleading, thus improving the quality of transcreation.
- Enhancing Interpretability: The system doesn't just improve translations; it also allows users to see why a particular translation was chosen by displaying the retrieved knowledge. This transparency helps users understand the MT process and build trust in the system.
Key Results
The paper demonstrates the effectiveness of Wiki-MT through its user interface (UI), which allows for side-by-side comparison of translations with and without knowledge graph integration. The example provided shows a clear improvement:
- Without Wiki-MT: "What is the plot of The Catcher in the Rye?" translated to Italian as "Qual è la trama di The Catcher in the Rye?" (incorrectly leaving the English title untranslated).
- With Wiki-MT: The same query is translated to "Qual è la trama di Il giovane Holden?" (correctly using the Italian transcreated title).
The system supports 11 languages and can be expanded to cover all languages supported by NLLB-200 and Wikidata.
Limitations and Future Work
- Limitations: The paper acknowledges that the MT system needs to be fine-tuned on a dataset of knowledge-seeking queries to effectively utilize the appended knowledge. The current implementation focuses on the top-3 entities, and further exploration of how to best leverage retrieved information could be beneficial.
- Future Work: The authors suggest two promising directions:
- Deeper Integration of Retrievers: Exploring ways to integrate retrievers more deeply into MT systems to incorporate richer semantics into the translation process.
- Improving Knowledge Graph Quality: Enhancing the quality of textual information within multilingual knowledge graphs themselves.
Summary updated