Difference between revisions of "Paralex"
From Knowitall
| Line 4: | Line 4: | ||
== Work Log == | == Work Log == | ||
| + | === Oct 25 === | ||
| + | Project on hold pending stable release of Paralex/Open-IE Demo. | ||
=== Oct 17 === | === Oct 17 === | ||
Shifting focus from this project to the Paralex Demo project for the time being - until a stable demo is up. | Shifting focus from this project to the Paralex Demo project for the time being - until a stable demo is up. | ||
Latest revision as of 21:58, 25 October 2013
Contents
Goals
- Paraphrase-Driven Open-Domain Question Answering
- Answer noisy user questions by generalizing from a question-paraphrase corpus.
Work Log
Oct 25
Project on hold pending stable release of Paralex/Open-IE Demo.
Oct 17
Shifting focus from this project to the Paralex Demo project for the time being - until a stable demo is up.
- Merged all of my code changes into the main branch, including new scoring features and training data.
- Discussed with Tony ideas for other sub-projects - next work will likely be on relation synonyms using PMI computed over the triplestore.
Oct 9
Discussed with Tony:
- Paraphrase module:
- New paraphrase template PMI scores are now based on more accurate template co-occurrence counts, and a minimum frequency of 50 was set for templates.
- Template language model (LM) trained based on n-grams from the lemmatized question corpus.
- Ideas for upcoming template classifier - Features for PMI, LM, pos-tag pattern, and leveraging the paraphrase corpus for unsupervised learning
- Answer Ranking
- AnswerDerivation ordering (baseline) ranking
- Strategies for training answer ranking function:
- Uniform sample of answers from representative questions?
- Use answers given in WikiAnswers for unsupervised learning?
Current Goals:
- Produce set of training (Q,A) pairs for hand-labeling from a uniform sample of answers over training questions.
- Explore feasibility of self-labeling data from answers given in WikiAnswers.
