Content uploaded by Peter Dekker
Author content
All content in this area was uploaded by Peter Dekker on Oct 26, 2018
Content may be subject to copyright.
Recognizing blends:
First experiments with
PYBOSSA
Peter Dekker
Tanneke Schoonheim
Instituut voor de Nederlandse Taal
enetCollect WG3&5 meeting Leiden, 24-25 October 2018
Introduction of INT
Instituut voor de Nederlandse Taal (Dutch Language Institute)
●Scholarly institute in the field of the Dutch language
●Central position in the Dutch-speaking world
●Developer, keeper and distributor of corpora, lexica, dictionaries
and grammars
●Provider of necessary building blocks of the study of Dutch
Introduction of INT
Current staff of INT:
●17 (computational and corpus) linguists,
lexicographers, terminologists,
4 linguistic assistants and trainees
●5 software engineers, 1 system administrator
●5 administration and communication
Introduction of INT
Current projects at INT:
●Contemporary and historical dictionaries and dictionary portals
●Contemporary and historical corpora and lexica
●Grammar portal, spelling database, terminology lists
●Infrastructure, tools and data for linguistic research (CLARIN)
Introduction of INT
Relatively new projects at INT:
Development and hosting of products for educational purposes, such as
●Bilingual dictionaries (New Greek, Portuguese, Estonian)
●Dutch Word Combinations
●Corpus Eenvoudig Nederlands (Corpus of Elementary Dutch)
Can crowdsourcing help us developing these and other educational products?
Research objective
Crowdsourcing:
●Task solved by public: answer unknown
●User details not important
Traditional (socio)linguistic research/survey:
●Answer of task known beforehand, in many cases
●User details important
Our research objective: combination
●PYBOSSA can be suitable, more solutions exist
PYBOSSA installation
●Hosted version (crowdcrafting.org) vs hosting on own server
○Own server: existing infrastructure at INT, more flexibility
●Clear installation guide on PYBOSSA website
●Complexity: PYBOSSA consists of multiple software packages
●Ansible script: recipe for reproducible installation
Experiments with blends: Data
Blend
●Compound of two words, where parts of the words are lost
●Signifies a new meaning, related to the words it consists of
Examples:
●glamping (glamour + camping)
●mup (millenial + yup)
Blends in English: Gries, S. T. (2004). Shouldn't it be breakfunch? A quantitative analysis of blend structure in English.
Linguistics, 639-668.
Experiments with blends: Data
Blends are part of two related projects at INT:
●Algemeen Nederlands Woordenboek (ANW; Dictionary of Contemporary Dutch)
●Neologism portal (upcoming)
Neologism Workflow at INT:
1. Data from newspapers and websites
2. Processed automatically, new words put aside
3. Lexicographer selects neologisms, creates entries in
a. neologism portal (all neologisms)
b. ANW dictionary (rooted neologisms only)
Experiments with blends: Crowd
Where did we find the crowd?
●Newsletter Instituut voor de Nederlandse Taal
●Congress Internationale Vereniging voor Neerlandistiek
Experiments with blends: Jobs
Can the crowd help in recognizing and analyzing blends?
Two jobs created in PYBOSSA:
●Blend recognition
○Recognize blends in a text
●Blend analysis
○Analyze the words a blend consists of
10 tasks per job
User interface design
Freedom in UI design: design using HTML and Javascript
PYBOSSA only loads and saves tasks from database
User details
Blends analysis
Blends recognition
Results: Age
Results: Gender
Results: Location
Results: Blends analysis
n = 326
Results: Blends analysis for preferendum
Analysis Frequency
referendum,
prefereren
154
referendum,
preferentie
60
referendum, pre 16
[Don’t know] 11
referendum,
preferent
8
●Multiple word forms (noun, verb) for prefer
●Multiple interpretations (at least when word is
presented without context):
○referendum + to prefer
○referendum + pre
Results: Blends recognition
n = 223
Results: Blends recognition for twittie
Recognized
blends
Frequency
twittie 122
twittie, fittie 56
fittie 16
twittie, tweet, fittie 5
[Do not know] 4
●twittie: twitter + fittie ‘fight’ (slang)
●fittie itself also occurred in text:
misinterpreted as blend
●More input fields (3) than real blends per task (1):
stimulates giving more blends
User feedback
●English language of PYBOSSA, while tasks are
about Dutch
●Too many buttons
●Task not always clear
●Make welcome page attractive
Experiences with PYBOSSA
Benefits
●Freedom when developing tasks
●Share tasks with other researchers
●Everything else (account system and loading/saving tasks) handled by PYBOSSA
●Quick answers from developers via bug tracker
Drawbacks
●No ready-made translation for all languages
●PYBOSSA not designed for asking user details
●When uploading large number of tasks, there is no clear end of job (you have to code that yourself)
●User cannot easily go back to a previous task
●User identification by IP address does not always work
Possible alternative for some purposes: Google Forms
Future experiments
●Neologisms and dialects
●User detail prediction as reward
Interesting issue:
●Is PyBossa better suited for these tasks than for instance Google forms?
Conclusion
Is crowdsourcing useful for the analysis of blends?
Yes, because it gives an insight in how blends are interpreted by non-linguists.
Is PYBOSSA useful for this kind of crowdsourcing?
Yes, powerful platform, with its own strengths and drawbacks.
