Generating Live Soccer-Match Commentary from Play Data
Authors: Yasufumi Taniguchi, Yukun Feng, Hiroya Takamura, Manabu Okumura7096-7103
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments on the play data of the English Premier League, provide a discussion on the result including generated commentaries. We conducted both automatic evaluation and human evaluation. For automatic evaluation, we used BLEU (Papineni et al. 2002) between the generated live commentaries and the gold-standard text. |
| Researcher Affiliation | Academia | Yasufumi Taniguchi,1 Yukun Feng,1 Hiroya Takamura,1,2 Manabu Okumura1 1Tokyo Institute of Technology 2National Institute of Advanced Industrial Science and Technology (AIST) |
| Pseudocode | No | The paper describes model architectures and equations but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Although the data itself has to be purchased, the code for preprocessing data and other related resources are available at https://github.com/yasufumy/placeholder reconstruction. |
| Open Datasets | No | We use play event data of soccer matches in the English Premier League3 for the 2015/16 season containing 380 soccer matches, provided by Opta Sports.4 The play data of each match consists of a sequence of events. The original dataset provided by Opta Sports contains 663,911 events and 26,340 commentaries. Although the data itself has to be purchased, the code for preprocessing data and other related resources are available at https://github.com/yasufumy/placeholder reconstruction. |
| Dataset Splits | Yes | We used 17,140 commentaries as our experimental data, out of which 13,662 were used as training data, 1,677 as development data, and 1,801 as test data. |
| Hardware Specification | No | The paper does not specify any particular hardware (e.g., CPU, GPU models, or cloud instances) used for running its experiments. |
| Software Dependencies | No | We used nltk7 for tokenizing the live commentary data, and the neural-network framework Chainer8 to implement our model. (No version numbers are provided for nltk or Chainer). |
| Experiment Setup | Yes | For the parameter optimization, we used Adam (Kingma and Ba 2014) and set the gradient clipping value (Pascanu, Mikolov, and Bengio 2012) to 5. We stopped the model training if the BLEU score on the development data became worse three times in a row, or the number of epochs reached 200. We set the dimension of the embedding vectors for the categorical values in events to 16, that of encoder s hidden state h to 200, and that of the decoder s word embedding vectors to 128. Each parameter is initialized using the Xavier initialization method (Glorot and Bengio 2010). During training, we truncate each commentary to 10 words. |