Dataset 2: Recommended Language Models
These word models were trained with a sentence start word of < s > , a sentence end word of < /s > , and an unknown word < unk > . The word vocabulary was the most frequent 64K words in the forum dataset that were also in a list of 330K known English words. All words are in lowercase. The character models are 12-gram models and were trained using interpolated Witten-Bell smoothing. The character model vocabulary consists of the lowercase letters a-z, apostrophe, < sp > ; for a space, < s > for sentence start, and < /s > for sentence end.
The perplexities in the above table are the average per-word or per-letter perplexity averaged on four evaluation test sets. The test sets were:
- Held out forum data from our mined collection of forum posts.
- Messages written by Enron employees on their Blackberry mobile devices, Enron mobile dataset
- Tweets written on a mobile device in the Summer of 2015.
- SMS messages from the NUS SMS corpus and the Mobile Forensics Text Message corpus
The above mixture models were trained on a total of 504M words of data: 126M words of forum data, 126M words from Twitter's streaming API between December 2010 and June 2012, 126M words of forum data from ICWSM 2011 Spinn3r dataset, and 126M words of blog data from the ICWSM 2009 Spinn3r dataset.
language models, text mining, text analysis
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Vertanen, Keith and Kristensson, Per Ola, "Dataset 2: Recommended Language Models" (2019). Mobile Text Dataset and Language Models. 2.