a1da4 / paper-survey

Summary of machine learning papers
32 stars 0 forks source link

Reading: A Tale of Two Laws of Semantic Change: Predicting Synonym Changes with Distributional Semantic Models #266

Open a1da4 opened 1 year ago

a1da4 commented 1 year ago

0. Paper

1. What is it?

2. What is amazing compared to previous works?

Previous research suffers from biases, so they release a dataset for analyzing the semantic shift of synonyms over time. Moreover, they propose two methods (unsupervised and supervised) for analyzing the semantic shift.

3. Where is the key to technologies and techniques?

3.1 Dataset

スクリーンショット 2023-06-03 21 49 55

They create a dataset using Fernald's English Synonyms and Antonyms and WordNet. They define synsets in WordNet as the "sense" of synonyms.

3.2 Unsupervised Methods

スクリーンショット 2023-06-03 21 53 20

In the prediction, they use a threshold (average DD or SD for all pair of words).

3.3 Supervised Methods

They use the above features to train the Logistic Regression model.

4. How did evaluate it?

スクリーンショット 2023-06-03 22 00 06

From this table, it is hard to predict synonyms or not correctly. Moreover, many methods predict most words as "Different (not synonym)".

5. Is there a discussion?

スクリーンショット 2023-06-03 22 03 00

They indicate that the low performance above table is due to a previous superordinate-subordinate relationship.

6. Which paper should read next?