Exploration of aunnotation strategies for entailment-based Automatic Short Answer Grading
View/ Open
Date
2023-06-30Author
Egaña Azpiazu, Aner
Metadata
Show full item recordAbstract
[EN] Recent work has shown that Automatic Short Answer Grading can effectively be
reformulated as a Textual Entailment problem. In this work we show that this
reformulation is also effective in zero-shot and few-shot settings, where we report
competent results close to state-of-the-art performance with the few-shot setting. More
importantly, we show that the annotation strategy can have significant impact on
performance. When annotating few examples, empirical results show that increasing the
variability on the question side, at cost of decreasing the amount of annotated answers
per question, is preferable than having the same number of annotated examples with less
questions and more answers. With this annotation strategy, using only the 10% of the full
training set our model levels with state-of-the-art systems in the SciEntsBank dataset.
Finally, experiments over SciEntsBank and Beetle domains show that the use of
out-of-domain annotated question-answer examples can be harmful, concluding that
task-aware fine-tuned models obtain significantly lower results compared to task-agnostic
general purpose inference models, at least with the domains employed for this work. [EU] Erantzun labur automatikoen sailkapenaren inguruan azken urteetan egindako ikerketek
atazaren birformulazio eraginkorra eraikitzea posible dela erakutsi dute, inferentzia
testualaren atazarako birformulazioa, bereziki. Gure lan honetan, birformulazioaren
eraginkortasuna erakusten da adibide gutxitako eszenarioetan (few-shot) eta adibide
gabeko eszenarioetan (zero-shot) ere bai. Are eta garrantzitsuago, atazarako adibideak
anotatzeko estrategiak modeloaren erredimenduan eragin nabarmena duela erakusten da.
Adibide gutxi batzuk idaztean, emaitza enpirikoek erakusten dute hobe dela galderaren
aldeko aldagarritasuna handitzea, galdera bakoitzeko idatzitako erantzun-kopurua
murriztearen kostuari dagokionez, galdera gutxiagorekin eta erantzun gehiagorekin
idatzitako adibide-kopuru bera izatea baino. Idazteko estrategia honi jarraituz,
entrenamendu osoko datu-basearen %10a erabiliz artearen egoerako sistemen
errendimenduaren parekoa da, SciEntsBank domeinuko datu-basean. Azkenik,
Beetle eta SciEntsBank domeinuen gainean aurrera eramandako esperimentuek
domeinuz kanpoko galdera-erantzun adibide bikoteek errendimendurako mingarriak izan
daitezkeela erakutsi dute, beste domeinu batetik ataza ezagutzen duten sistemek ataza
ezagutzen ez dutenak baino emaitza apalagoak emateko joera dutela ondorioztatuz,
aztertutako domeinuetan behintzat.