Introduction
I recently discovered some incredible research around modeling how the readers might comprehend a text. A recent paper titled "Automated Model of Comprehension V2.0" (Corlatescu et. al) outlined a computational model of reading comprehension. This and other such models are based on cognitive science theories that describe how the brain functions in various contexts. For example, during reading (and possibly writing), the short-term memory system may be involved, and the computational models would somehow try to emulate memory. This is how Corlatescu et. al give the background behind their model:
"How well a reader understands text or discourse depends on many factors, including individual differences such as reading skill, prior knowledge of the domain or world, motivation, and goals. Comprehension also depends on the nature of the text – the difficulties imposed by the words in the text, the complexity of the syntax, and the flow of the ideas, or cohesion.
Cohesion between ideas can emerge from the overlap between explicit words (e.g., nouns, verbs), implied words (anaphor), semantically related words, semantically related ideas, and the underlying parts of speech (i.e., parts of speech, syntactic overlap). When there is greater overlap, the text is easier to understand. Cohesion gaps, by contrast, require inferences to make connections between the ideas. If the reader has little knowledge of the domain or the world, low cohesion text impedes comprehension." (Corlatescu et. al)
Automated Model of Comprehension (AMoC)
The paper “Towards an Automated Model of Comprehension (AMoC)” (Dascalu et. al) describes the model in the aforementioned paper, along with two other models, named Construction Integration Model and the Landscape Model. All of these models take a text passage as input and build a hypothetical model of what the text comprehension may involve. The AMoC model also has some tuning parameters. Here is how Corlatescu et. al describes the AMoC model:
"The automated model of comprehension (AMoC) simulates the mental representation constructed by hypothetical readers, by building syntactic and semantic relations between words, coupled with inferences of related concepts that rely on various semantic models." (Corlatescu et. al)
Example
Let us try an example of AMoC! We are going to input a text passage into the AMoC model and see what reading comprehension model it prints out for us. You can do it yourself here: http://readerbench.com/demo/amoc (it will be a little slow but it works).
Sentence: The ministry of health announced the decision to change the gap from 6-8 weeks to 12-16 weeks on May 13, at a time when supplies of the shot were falling short of demand and infections were surging across the country.
Result
This is how Corlatescu et. al describe we can interpret the visualization: “The inner circle depicts in blue text-based information that is still active, while the outer circle contains semantically inferred concepts in red and grayed out inactive concepts.” I suspect that this length of my input has created a model that seems a little odd. But feel free to try some other examples on your own, and see what you get!
This paper was published at the (ongoing) 2021 Artificial Intelligence in Education conference.
Thank you for reading and have a great day!
Comentarios