Learning Better Name Translation for Cross-Lingual Wikification
Authors: Chen-Tse Tsai, Dan Roth
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comparing to 6 other approaches on 9 languages, we show that the proposed model outperforms others not only on the transliteration metric, but also on the ability to generate target English titles for a cross-lingual wikifier. Consequently, as we show, it improves the end-to-end performance of a cross-lingual wikifier on the TAC 2016 EDL dataset. |
| Researcher Affiliation | Collaboration | Chen-Tse Tsai Bloomberg LP New York, NY ctsai54@bloomberg.net Dan Roth University of Pennsylvania Philadelphia, PA danroth@seas.upenn.edu |
| Pseudocode | No | The paper describes mathematical models and update rules, but does not include a block explicitly labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | No | The paper mentions augmenting a base system available at 'https://github.com/CogComp/cross-lingual-wikifier', but does not explicitly state that the code for their proposed model is open-sourced or provide a direct link to it. |
| Open Datasets | No | We create training, development, and test title pairs from the inter-language links in Wikipedia. For a test language L, we take all the titles in L s Wikipedia which have a link pointing to the corresponding English page, and then use Free Base types to classify them into one of the three entity types: person, location, and organization, or discard a title if it is not of any of these types. The paper describes how they created their dataset but does not provide specific access information (link, DOI, repository) for the processed dataset used in their experiments. |
| Dataset Splits | Yes | For each entity type, we take at most 10k pairs for training and 5k pairs for both development and test. The numbers of title pairs for each language are shown in the column #Title Pairs of Table 2. Table 2 provides explicit numbers for Train, Dev, and Test sets for each language and entity type. |
| Hardware Specification | No | The paper does not specify the hardware used for running the experiments. |
| Software Dependencies | No | The paper mentions various models and tools such as 'Direc TL+', 'Sequitur', 'P&R', 'JANUS', 'NMT-bpe', 'NMT-char', 'm2m-aligner', and 'word alignment (Dyer, Chahuneau, and Smith 2013)', but does not provide specific version numbers for any of these software dependencies. |
| Experiment Setup | Yes | We use 500 dimensional embeddings and 100 training epochs. The top 10 frequent foreign words are usually selected. The language model is trained on all articles in the English Wikipedia. For each entity type, we take at most 10k pairs for training and 5k pairs for both development and test. We query the first dictionary by the entire mention string to retrieve the top 30 titles. If there are less than 30 titles, we then query the second dictionary by each word in the mention. The third dictionary is used in a similar way if the total number of candidates is still less than 30. |