On the rules given to lms programs using artificial intelligence with the help of natural language processing

Yüklə 462,45 Kb.
Pdf görüntüsü
ölçüsü462,45 Kb.
1   2   3   4   5

The Designed System
The designed system is explained in general, as well as its each component. 
Mazidi and Tarau (2016) state that natural language understanding (NLU) is a 
missing piece of the puzzle in question generation. So, in a similar vein, the main 
focus of this study is NLU. It is achieved by exploiting the semantic roles of words.
The designed system focuses both on deep and shallow questions which have 
mentioned. As shallow questions, the designed system can generate the following 
question sentences:
• What...?
• Who...?
• Where...?
• How many...?
• When...?
As deep questions, the designed 
system can generate:
• Why...?
• How...?
• How would you describe...?
• Indicate characteristics of...?
• For what purpose...? 
Flowchart of the designed system 

Conclusion and Future work
This thesis presented a rule based automatic question generation system that 
focuses on both question generations from sentences and paragraphs. Especially, 
with respect to METEOR metric, the designed system significantly outperforms all 
other systems in automatic evaluation stage. Banerjee et al. (2005) demonstrated that 
METEOR has significantly enhanced correlation with human evaluators. So, our 
results confirm that statement by performing human evaluation study. In conclusion, 
the designed system significantly outperforms all other systems in human evaluation 
study by generating the most natural (human-like) questions.
For deciding between who and what questions, we proposed solution. This 
problem is one of the lexical challenges that we have stated in. Our results in Table 
shows that, with 4.262 correctness score, we correctly differentiate between who and 
what questions. Also, for another lexical challenge, non-compositionality that is 
stated, we proposed solution. Our predefined dictionary does not cover all idioms. 
Also, some types of idioms cannot be covered with predefined dictionary. This issue 
will be explored in the future work.
Currently, our templates do not achieve the best performance across all 
question categories. If we look at Table, S-V-number and S-V-ARGM-MNR (how) 
type of questions has a low correctness score. In addition, in order to improve the 
performance of paragraph based questions in all templates, we need to investigate 
how to better use the paragraph-level information. This is one of the discourse 
challenges that we have mentioned. Information conveyed from one sentence to 
other is a problematic issue. So, we leave this issue to future work. Finally, some 
templates fit better with some topics than others. For instance, S-V-attr and S-V-
oprd templates that is stated, works better with noun phrases that are suitable with 
descriptive questions. For definition questions, other techniques need to be explored 
in the future work.
Also, adapting the designed system to Uzbek language would not be easy due 
to lack of syntactic and semantic parsers. Without high-performance parsers

adapting predefined rules into Uzbek language would not give a similar 

Yüklə 462,45 Kb.

Dostları ilə paylaş:
1   2   3   4   5

Verilənlər bazası müəlliflik hüququ ilə müdafiə olunur ©www.genderi.org 2024
rəhbərliyinə müraciət

    Ana səhifə