दल
DIL: Dialectic Intercultural Lexicon
POEM GENERATOR
2024
Dreams by दल
Whispers from Multiverse by दल
Poetry in its essence consists of an interplay between imagery, language, rhythm, rhetoric and metaphoric nodes of expressing cultural nuances into experiential forms. Our prototyped AI system ‘दल- DIL: Dialectic Intercultural Lexicon’ projects this notion as a polyglottal-model system which enables its user to create poems with diverse languages. This system is capable of creating poems in any language when relevant datasets are provided. Additionally, the system requires the translation of every poem that is generated in a previously chosen language, for instance, English, which can be achieved using ChatGPT. The prototype presented in this iteration consists of poems, incorporating four languages: Dutch, Polish, Hindi and Spanish using दल. The poems generated are stemming from data sets consisting of songs about love in each language. This prototype is inspired by recognising the multicultural feature of our group as it consists of people representing four different oral traditions, as those among us communicate daily in Dutch, Hindi, Polish and Spanish. In order to celebrate this diversity, we decided to create an generative AI system that aimed at creating poetry that is polyglottal, i. e. consists of vocabulary of multiple languages. As we were addressing the sublime medium of poetry, we agreed on focusing on the appropriate theme of love.
​
In order to achieve a task of creating poems full of non-trivial relationships between each of the languages, we built a markov chain (n-gram) algorithm (markovchainngram.py). This model generated text in an autocomplete manner, meaning that it generated an n amount of words based on the last word in the sentence. As a result of this approach, our polyglottal poem generator found its way of combining data sets based on common words between the languages.
​
Database: Our designed dataset consisted of ten songs about love in all four languages, Dutch, Polish, Hindi and Spanish. Each group member selected poems from their own language(s). It is worth noting that Hindi, unlike Dutch, Polish and Spanish that use latin alphabet, is written in its own script called devanagari (alphabet used for Sanskrit, Hindi and other Indian languages).
​
​Stopping criterion: The stopping criterion for the poems is based on where the songs in the dataset would naturally end. There is an end character programmatically added at the end of the songs in the dataset that can naturally show up at the end of a markov chain. Here it will function as the end point of the poem. ​
Presentation of Unfinished- Chronicles by दल. The text is translated to English (using ChatGPT) and the recitation is based on the original languages the peom was generated in. Tip: Please experience with the sound on, by clicking the beamed music note on bottom-right of the video.
Presentation: The consequence of our poems being polyglottal required the reader to be well versed in more than one of the used languages. Having acknowledged this obstacle against fully enjoying the generated poetry by the broad audience, we understood the task of presenting the translation as the crucial part of the poems’ presentation. Our task was to find a medium that conveys both the eccentrical appeal of a polyglottal message and at the same time clear and transparent meaning (to the extent we naïvelt trust the institution of translation). Thus we decided on the multimodality within the way the poems are presented. As the original sound of the poems occupy the ever-changing auditory realm, the visual realm solidly persists on the screen, as the automatic typer unfolds the poem’s lyrics on the screen. We used ChatGPT in the role of the AI output generator in the sense that the language model provides a coherent English translation of each of the poems. Thus, the medium of our AI-generated poetry can be the very machine that those poems were generated on and our final output of the assignment takes the form of a Python script and user’s terminal window.
Text: The translated text of the poem appears letter-by-letter in the terminal window. The “typewriter” style of the visuals are achieved using a command-line tool “clicklick” and Python’s “keyboard” module.
Recitation: The original text of the poem is recited by OS-native speech assistants according to the language of the current passus of the poem.
Sound: Typewriting each of the poems is accompanied by sound effects using various digital instruments supplied within the Python’s SCAMP module. Those sounds aim to entertain the otherwise quiet process of typewriting and both worlds of the auditory and visual realms of the presentation.
​
Evaluation of Creativity: The creativity of our AI model, ‘दल- DIL: Dialectic Intercultural Lexicon’ can be evaluated using the Lovelace 2.0 test. Our system दल generated two creative outputs ie. auditory output p (polyglottal poems) and visual output (typed p’ based on a provided theme/genre l (in our case: the set of 4 languages) by us, the human evaluators h, given the constraints s being the stopping criterion of the poems under the supervision of the users u. Considering the aforementioned variables, and conditions, our system suffices to the assessment of the Lovelace test as दल is generating p and p’ as a result of l and s by h for the experience of u. Each iteration of p and p’ is novel in essence due to the condition s. दल ceases to generate p and p’ when it encounters a word that is at the end of a poem in the provided dataset l. Therefore, the results generated by दल are novel to the experience of has well as for u. Another intriguing element about दल is that p and p’ are also producing uniquesentences/phrases inspired from the dataset: as witnessed in the poem named Echos, in its fourth stanza - देर से van (der se van) which is a combination of Spanish and Hindi words that दल identified to be similar in nature in both languages to form this phrase, meaning ‘’lately they go…’’.
Within the framework of computational creativity, our AI system presented the features of originality and is solution/result based. दल is designed to work towards producing an end result, with the intention to produce something that did not exist before. An example of this can be seen in the poem ‘Whispers of Multiverse’ (as shown above) consisting of the phrase ‘’la escoba que en tu vida barra la tristeza’’ which translates to ‘’let the broom sweep away the sadness in your life’’. This AI model also operates as an independent entity as its autonomy allows it to compose sentences, stanzas, phrases using individual words from the dataset, giving it more freedom to produce its own polyglottal language, in essence challenging the cultural and linguistic norms. The presentation AI model consisting of both auditory and visual outputs adds another layer of engagement for the user. The auditory output, personifies दल, as a ‘being’ or ‘artificial creature’ reciting these poems to the user. The simultaneous typing of the text in the form of translation of the polyglottal poems in English, invites the user to engage with the system’s multilingual expression of love.
​
This collaborative work by Artur Dobija, Pragya Jain and Rick Heemskerk, was developed in the context of Computational Creativity course by Rob Sauders.
PRAGYA_JAIN_2025