Abstract—This paper examines text similarity approach based on Google n-gram dataset. Google Tri-grams Measure (GTM) is an unsupervised text similarity measure. The paper investigates the sentence similarity of GTM which in turn reveals the approach’s pitfalls. We also compared GTM’s sentence similarity measures on Li-30 sentence pairs, Microsoft Research Paraphrase Corpus paraphrase, Kaggle Quora Question Pairs competition’s dataset respectively against human judgement. Other sentence similarity measures are compared against GTM. We discovered GTM sentence similarity has a lot of weight on overlapped words count. However, despite the weakness, it still outperformed other replicated sentence similarity measures.
Index Terms—Google trigrams, pitfalls, sentence similarity, text similarity, trigrams, unsupervised, word similarity.
The authors are with Faculty Computer Science and Information Technology, Universiti Malaysia Sarawak, Malaysia (e-mail: 15020282@siswa.unimas.my, chbong@unimas.my, nklee@unimas.my).
[PDF]
Cite: Wong Lin Juan Linda, Chih How Bong, and Nung Kion Lee, "Re-examining Google Tri-grams Measure (GTM) Sentence Similarity," International Journal of Knowledge Engineering vol. 3, no. 2, pp. 80-85, 2017.
