GTAP Resources: Resource Display
| GTAP Resource #7497 |
|---|
|
"Enhancing the readability of HS descriptions: Automatic Text Summarisation using LLMs" by Ngavozafy, Antonia, Youssef Mamlouk and Maria C. Latorre Abstract This paper presents an automatic text summarisation method to generate a lexically simplified and shorter version of product descriptions in the Harmonised System (HS) using Large Language Models (LLMs). The results demonstrated that GPT-4 is effective in enhancing the readability of HS descriptions. On a scale of zero to five, the model was observed to reduce their complexity by 1.09 points when evaluated by machine-emulated high school student having an intermediate level of English and by 0.53 points when assessed by human trade professionals. Moreover, the summarisation has markedly enhanced the readability of highly complex descriptions, with a score of 5, which accounted for 14.8% of the dataset after the summarisation process instead of 36.7% initially. The use of semantic textual similarity as an automatic evaluation metric, coupled with human-perceived informativeness, has ascertained that the summaries are semantically and factually consistent with the original HS descriptions. Basic BERT and PEGASUS are limited in their ability to generate the desired lexically simplified and summarised version of the HS descriptions. On the one hand, BERT generates semantically inconsistent summaries even after increasing the candidate pools, improving complexity detection and relaxing filtering thresholds. For PEGASUS, paraphrasing gives the best results in terms of semantic preservation, but the structure of the output often does not match the original input. Consequently, both BERT and PEGASUS may require more advanced parameter adjustment and fine-tuning to increase their efficiency. The lexically simplified and summarised HS descriptions resulting from our ATS procedure, designated as the SIMPLEX, can be utilised in a multitude of trade compliance software and web-applications, such as a new generation of HS Search Engine, thereby expediting the reading and processing of the product descriptions or serving as a training dataset of a similarity search engine. |
| Resource Details (Export Citation) | GTAP Keywords | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
- Other data bases and data issues - Technological change - Multilateral trade negotiations - Advances in quantitative methods - Software and modeling tools - Global |
| Attachments |
|---|
|
If you have trouble accessing any of the attachments below due to disability, please contact the authors listed above.
Public Access  Paper (1.1 MB) Replicated: 0 time(s)Restricted Access No documents have been attached. Special Instructions Working Paper - Not for Publication |
| Comments (0 posted) |
|---|
|
You must log in before entering comments.
No comments have been posted. |
Last Modified: 9/15/2023 2:05:45 PM
