Week 16 Sem 2

Conclusions Observation & Future Work...

We need for larger databases to make things more interesting. More importantly we have identified a certain key issue to address.

Slow WSD processing.
When the document contains a lot of line the WSD part takes a lot of time to process. Although BlogWall was lucky to have small documents (SMS messages) to process on run time this might affect other project when incorporated. This is because of the highly complicated measurement of sense relatedness along with the addition of the dictionary features. More categories will help make this faster. Threading will also help.

New vs Old System Performance

New solution works well especially when term frequency of a term is low. And similarly the old systems work well, sometimes even better than the new system when there is sufficient term frequency.

 

8.5      Analysis and Discussion

Since the whole point of judging a poem is subjective, the need for user study is quite stark. My point of view is simply just mine. Hence we planned to conduct some user studies.  And here are the results. It was conducted in a fair environment, where 15 random SMSes were used to generate results in 3 alternative system 1- the old one, 2- the new one (based on WSD) and the third a system based on the old one but has learning incorporated (SVM based). 36 users were asked to rate the relevance & meaningfulness of the poems corresponding to the inputs. They were also asked to give a satisfaction rating and here are the results.

These are extremely good results with an increase of 38 % in user satisfaction for the new system. Also 4 out of 5 SMSes receive their best output through the new system. This is in line with what we predicted earlier and hence this procedure was worth the effort.

Chapter 9: CONCLUSION

9.1      Project Conclusion

The BlogWall Poetry Mix-up is a novel mobile artistic media application which promotes self-expression and public communication, combining media art and poetry. While text messages are used as the interface now, the poetry generation algorithm will work for any text. The application combines data mining, part of speech tagging and emotional analysis to generate poetry. It is an attempt to recreate social communication among the youth by drawing from the phenomena of “mixing” or “mash-up” and applying these ideas to poetry. The novel interface to poetry as well as wide usage of text messages among younger generation would make this application very appealing. With this effort we hope to create new form of text message art as well as attractant younger generation to literary works such as poetry.

 

Every one of us has some level of artistic or poetic ability. However, not everyone is able to create a poem. Especially, the younger generation may not have the necessary background or the knowledge to do so. The BlogWall is a technique to bridge this gap. Text messages provide the ideal basics because the technology is familiar. The interface, using a mobile phone is also within easy reach.  The BlogWall allows each individual to create their own custom and unique work of personal art, by using existing work. Thus, a very personal and unique poetry mixer can be created depending on the input text.

9.2      Components Conclusion & Related Work

From the experiments we have sufficient evidence to suggest that the methods employed are efficient enough. But the truth of the matter is that there are new measures waiting to be discovered. Especially the Word Sense Disambiguation layer which could benefit from a better unification of WordNet & an open source dictionary along with improved semantic relatedness measures including improving detecting of gloss overlaps etc. Also new ways to measure semantic relatedness faster might help a lot. Also the topic summarization could have more functions to calculate salience apart from semantic relatedness. As far as the recombination algorithm is concerned, I believe it is just a start. Currently it possesses basic ability to remix poem lines. Maybe someday it can add / edit / remove words from a poem line to better suit the input message and perhaps include rhythm & rhyme (sound aspect) of the poem in its calculations.

9.3      Individual Component Applications

As mentioned in the thesis the various components built could be used for various purposes. By making the Word Sense Disambiguation layer and the topic summarizing layer totally independent modules we have enabled its re use in several other projects especially those that require Natural Language Processing or even Natural Language Generation. One use that we are looking at specifically is to create remixed books – specifically technical ones.

 

Another aspect that could be explored is the application of these tools on the world’s largest database – the internet. This can potentially lead us to great applications.


 

Chapter 10:           FUTURE WORK

Instead of just using text to communicate the output, the interface could be much richer. Exploring the use of richer media such as sounds, music and pictures could also be interesting.

Furthermore, to encourage user participation, additional ways of interaction could be provided to the user. Allowing users to type their inputs directly would be a start. However, to take it further, the poetry mixer could, for example, be ported to the web. This would make it a very social application where people could export and share the works they have created with others.

 

Another interesting possibility, which unfortunately could not be explored in this thesis, is to limit the database of poetry to say a particular genre or author. In this case, the poetry mixer will essentially become a ‘specialized’ application which generates new poems but from a specific set. Examples could include a “Shakespearean poetry generator” or “Limerick generator”. This could increase the appeal of the application, and also potentially increase the effectiveness of coming up with a poem that is relevant and entertaining.

 

Another possibility is to explore the possibility of adding, editing and removing words in a line to better suit the input message something like what Manurung suggests (17).  It could also have a layer that checks for rhyme and rhythm aspects of the poem in the recombination algorithm part. Now that the Word Sense Disambiguation Layer is up with the other tools this should not be too far from actualization. Appropriate machine learning may also be added to combine the systems to produce the best results when conditions favor either of those solutions

Comments