in Basque using Natural Language Generation Manex Agirrezabal Euskal Herriko Unibertsitatea / University of the Basque Country (UPV/EHU) [email protected] October 13th, 2014 1/23
verses unique: phonetics and meaning Phonetics: We can get the metrical structure for verses in Basque and English We can generate verses that follow a specified meter in Basque Meaning: We could use topic modelling to get the possible topic of the verse But, how to generate a verse about a specific topic? 5/23
so far: Basque verse corpora. TXT → TEI P5 Semantic searcher for Basque verses (mg4j IR system) Analysis: Metrical analysis for verses in Basque (Agirrezabal et al., 2012) 6/23
so far: Basque verse corpora. TXT → TEI P5 Semantic searcher for Basque verses (mg4j IR system) Analysis: Metrical analysis for verses in Basque (Agirrezabal et al., 2012) Statistical analysis of Basque verses (Agirrezabal et al., 2013a) 6/23
so far: Basque verse corpora. TXT → TEI P5 Semantic searcher for Basque verses (mg4j IR system) Analysis: Metrical analysis for verses in Basque (Agirrezabal et al., 2012) Statistical analysis of Basque verses (Agirrezabal et al., 2013a) ZeuScansion: A tool for scansion of English poetry (Agirrezabal et al., 2013b) 6/23
so far: Basque verse corpora. TXT → TEI P5 Semantic searcher for Basque verses (mg4j IR system) Analysis: Metrical analysis for verses in Basque (Agirrezabal et al., 2012) Statistical analysis of Basque verses (Agirrezabal et al., 2013a) ZeuScansion: A tool for scansion of English poetry (Agirrezabal et al., 2013b) Three approaches for stress assignment to out-of-vocabulary words in English (Agirrezabal et al., 2014) 6/23
so far: Basque verse corpora. TXT → TEI P5 Semantic searcher for Basque verses (mg4j IR system) Analysis: Metrical analysis for verses in Basque (Agirrezabal et al., 2012) Statistical analysis of Basque verses (Agirrezabal et al., 2013a) ZeuScansion: A tool for scansion of English poetry (Agirrezabal et al., 2013b) Three approaches for stress assignment to out-of-vocabulary words in English (Agirrezabal et al., 2014) Generation: Surface realization using POS-tag templates from Basque verse corpora (Agirrezabal et al., 2013c) 6/23
in the Basque Country It’s a long standing tradition Championships, get-togethers and performances quite common Basque verse-making championship Celebrated every 4 years 7/23
in the Basque Country It’s a long standing tradition Championships, get-togethers and performances quite common Basque verse-making championship Celebrated every 4 years Congregates around 15.000 people 7/23
typical task: Given a topic, verse-makers have to sing three verses extempore They compose the verses in a limited time Each verse has constraints, such as: Fixed number of syllables Fixed number of lines Some lines have to rhyme 8/23
typical task: Given a topic, verse-makers have to sing three verses extempore They compose the verses in a limited time Each verse has constraints, such as: Fixed number of syllables Fixed number of lines Some lines have to rhyme Then, they sing the verse 8/23
typical task: Given a topic, verse-makers have to sing three verses extempore They compose the verses in a limited time Each verse has constraints, such as: Fixed number of syllables Fixed number of lines Some lines have to rhyme Then, they sing the verse ...but there are more tasks 8/23
is spoken along the Basque Country Approximately 700.000 speakers Morphologically rich language (multiple declension cases) There is a standardized form, but also 4-5 dialects 9/23
to follow the architecture for Natural Language Generation proposed in (Reiter et al., 2000). Six steps for the generation process: Content determination Document structuring Lexicalization Aggregation Referring expression generation Surface realization 12/23
One topic Output: Set of possible content to be included in the verse 1 Lexical resources (Some dictionaries or WordNet) 2 Machine learning (Content word mappings from topic to verses) 3 Topic modelling (Closest words according to some distance, like the cosine distance) 13/23
Set of content to be included in the verse Output: Content to be included and structure of the verse (Plan of the actions or chunks) Use existing AI planning systems (verbs as states) Take vector representations of words and produce similar verb paths 14/23
and plan of the verse Output: Specific words for each action or chunk in the plan Although we have not done too much work on this topic, we intend to use the results of the statistical verse analysis that we did last year. The most used rhyme patterns The use of informal language The use of borrowed words from Spanish (not Basque words) The use of standardized language 15/23
Structured content and specific words to be included in each chunk Output: Actual text for each action We are going to add each chunk’s information in existing POS-tag templates. 17/23
to create a verse-singing robot by incorporating some simple verse-making algorithms (that included strophe combining or IR) with a modified TTS system (adapted for singing) and a robot. 19/23
to create a verse-singing robot by incorporating some simple verse-making algorithms (that included strophe combining or IR) with a modified TTS system (adapted for singing) and a robot. This was a collaboration between our research group, which works on NLP, and other two research groups (Robotics lab and Signal processing lab). All of them from the University of the Basque Country. 19/23
in Basque using Natural Language Generation Manex Agirrezabal Euskal Herriko Unibertsitatea / University of the Basque Country (UPV/EHU) [email protected] October 13th, 2014 22/23