Difference between revisions of "Settles, CALC 2010"
| (5 intermediate revisions by the same user not shown) | |||
| Line 5: | Line 5: | ||
== Online version == | == Online version == | ||
| − | [http://www.cs.cmu.edu/~bsettles/pub/ | + | [http://www.cs.cmu.edu/~bsettles/pub/settles.calc10.pdf From Author's Webpage] |
== Summary == | == Summary == | ||
| − | This [[Category::Paper| paper]] describes two natural language processing based computational creativity tools to assist songwriters, originally deployed to aid songwriters participating in FAWM, a social challenge website devoted to writing songs in short period of time. The two tools described are Titular, a text synthesis algorithm for generating new song titles, and LyriCloud, which is a word-level explorer that allow users interactively select words and receive suggestions for lyrical words. | + | This [[Category::Paper| paper]] describes two natural language processing based computational creativity tools to assist songwriters, originally deployed to aid songwriters participating in FAWM (February Album Writing Month), a social challenge website devoted to writing songs in short period of time. The two tools described are Titular, a [[AddressesProblem::Natural Language Generation|text synthesis]] algorithm for generating new song titles, and LyriCloud, which is a word-level explorer that allow users interactively select words and receive suggestions for lyrical words. |
'''Titular''' | '''Titular''' | ||
| Line 15: | Line 15: | ||
Available online here: [http://muse.fawm.org/titular] | Available online here: [http://muse.fawm.org/titular] | ||
| − | This tool makes use of a template-based approach to synthesis titles. Table 1 shows some sample templates that were induced by Titual. As opposed to using hand-crafted templates, Titular learns using the following: | + | This tool makes use of a [[UsesMethod::Template Based Natural Language Generation|template-based]] approach to synthesis titles. Table 1 shows some sample templates that were induced by Titual. As opposed to using hand-crafted templates, Titular learns using the following: |
* Titles in training corpus are tokenized and POS tagged | * Titles in training corpus are tokenized and POS tagged | ||
* Open word classes are substitued with their POS tag, and close words remain in tact | * Open word classes are substitued with their POS tag, and close words remain in tact | ||
| Line 26: | Line 26: | ||
Available online here: [http://muse.fawm.org/lyricloud] | Available online here: [http://muse.fawm.org/lyricloud] | ||
| − | Figure 2 shows some example output by LyriCloud. | + | The original method of generating the tag clouds involved [[UsesMethod::Latent Dirichlet Allocation]], but was rejected during pilot testing because they failed to see how the generated words were connected to the seeds and were often high frequency words that were not very creative. Instead the author opted for a variant of a [[UsesMethod::Log Tempered TF-IDF|log-tempered tf-idf]] (from information retrieval literature). Figure 2 shows some example output by the current LyriCloud system. |
[[File:settles-calc-2010-2.png]] | [[File:settles-calc-2010-2.png]] | ||
== Related papers == | == Related papers == | ||
Latest revision as of 11:50, 31 March 2011
Citation
Computational Creativity Tools for Songwriters. Burr Settles. In Proceedings of the NAACL HLT 2010 Second Workshop on Computational Approaches to Linguistic Creativity, 2010.
Online version
Summary
This paper describes two natural language processing based computational creativity tools to assist songwriters, originally deployed to aid songwriters participating in FAWM (February Album Writing Month), a social challenge website devoted to writing songs in short period of time. The two tools described are Titular, a text synthesis algorithm for generating new song titles, and LyriCloud, which is a word-level explorer that allow users interactively select words and receive suggestions for lyrical words.
Titular
Available online here: [1]
This tool makes use of a template-based approach to synthesis titles. Table 1 shows some sample templates that were induced by Titual. As opposed to using hand-crafted templates, Titular learns using the following:
- Titles in training corpus are tokenized and POS tagged
- Open word classes are substitued with their POS tag, and close words remain in tact
- Thresholds are set to on templates to remove junk/infrequent templates
LyriCloud
Available online here: [2]
The original method of generating the tag clouds involved Latent Dirichlet Allocation, but was rejected during pilot testing because they failed to see how the generated words were connected to the seeds and were often high frequency words that were not very creative. Instead the author opted for a variant of a log-tempered tf-idf (from information retrieval literature). Figure 2 shows some example output by the current LyriCloud system.
