Sentence Embeddings
Inspired by Thai Sentence Vector Benchmark, we decided to embark on the journey of training Indonesian sentence embedding models!
Evaluation
Semantic Textual Similarity
We believe that a synthetic baseline is better than no baseline. Therefore, we followed approached done in the Thai Sentence Vector Benchmark project and translated the STS-B dev and test set to Indonesian via Google Translate API. This dataset will be used to evaluate our model's Spearman correlation score on the translated test set.
You can find the translated dataset on 🤗 HuggingFace Hub.
Retrieval
To evaluate our models' capability to perform retrieval tasks, we evaluate them on Indonesian subsets of MIRACL and TyDiQA datasets. In both datasets, the model's ability to retrieve relevant documents given a query is tested. We employ R@1 (top-1 accuracy), MRR@10, and nDCG@10 metrics to measure our model's performance.
Classification
For text classification, we will be doing emotion classification and sentiment analysis on the EmoT and SmSA subsets of IndoNLU, respectively. To do so, we will be doing the same approach as Thai Sentence Vector Benchmark and simply fit a Linear SVC on sentence representations of our texts with their corresponding labels. Thus, unlike conventional fine-tuning method where the backbone model is also updated, the Sentence Transformer stays frozen in our case; with only the classification head being trained.
Further, we will evaluate our models using the official MTEB code that contains two Indonesian classification subtasks: MassiveIntentClassification (id)
and MassiveScenarioClassification (id)
.
Pair Classification
We followed MTEB's PairClassification evaluation procedure for pair classification. Specifically for zero-shot natural language inference tasks, all neutral pairs are dropped, while contradictions and entailments are re-mapped as 0
s and 1
s. The maximum average precision (AP) score is found by finding the best threshold value.
We leverage the IndoNLI dataset's two test subsets: test_lay
and test_expert
.
Methods
(Unsupervised) SimCSE
We followed SimCSE: Simple Contrastive Learning of Sentence Embeddings and trained a sentence embedding model in an unsupervised fashion. Unsupervised SimCSE allows us to leverage an unsupervised corpus -- which are plenty -- and with different dropout masks in the encoder, contrastively learn sentence representations. This is parallel with the situation that there is a lack of supervised Indonesian sentence similarity datasets, hence SimCSE is a natural first move into this field. We used the Sentence Transformer implementation of SimCSE.
ConGen
Like SimCSE, ConGen: Unsupervised Control and Generalization Distillation For Sentence Representation is another unsupervised technique to train a sentence embedding model. Since it is in-part a distillation method, ConGen relies on a teacher model which will then be distilled to a student model. The original paper proposes back-translation as the best data augmentation technique. However, due to the lack of resources, we implemented word deletion, which was found to be on-par with back-translation despite being trivial. We used the official ConGen implementation which was written on top of the Sentence Transformers library.
SCT
SCT: An Efficient Self-Supervised Cross-View Training For Sentence Embedding is another unsupervised technique to train a sentence embedding model. It is very similar to ConGen in its knowledge distillation methodology, but also supports self-supervised training procedure without a teacher model. The original paper proposes back-translation as its data augmentation technique, but we implemented single-word deletion and found it to perform better than our backtranslated corpus. We used the official SCT implementation which was written on top of the Sentence Transformers library.
Pretrained Models
Deprecated Models
Model | #params | Base/Student Model | Teacher Model | Train Dataset | Supervised |
---|---|---|---|---|---|
SimCSE-IndoBERT Lite Base | 12M | IndoBERT Lite Base | N/A | Wikipedia | |
SimCSE-IndoRoBERTa Base | 125M | IndoRoBERTa Base | N/A | Wikipedia | |
S-IndoBERT Base mMARCO | 125M | IndoBERT Base | N/A | mMARCO | ✅ |
all-IndoBERT Base p2 | 125M | IndoBERT Base p2 | N/A | See: README | ✅ |
Results
Semantic Textual Similarity
Machine Translated Indonesian STS-B
Model | Spearman's Correlation (%) ↑ |
---|---|
SimCSE-IndoBERT Base | 70.13 |
ConGen-IndoBERT Lite Base | 79.97 |
ConGen-IndoBERT Base | 80.47 |
ConGen-SimCSE-IndoBERT Base | 81.16 |
ConGen-Indo-e5 Small | 80.94 |
SCT-IndoBERT Base | 74.56 |
all-IndoBERT Base | 73.84 |
all-IndoBERT Base-v2 | 76.03 |
all-Indo-e5 Small-v2 | 79.57 |
all-Indo-e5 Small-v3 | 79.95 |
distiluse-base-multilingual-cased-v2 | 75.08 |
paraphrase-multilingual-mpnet-base-v2 | 83.83 |
multilingual-e5-small | 78.89 |
multilingual-e5-base | 79.72 |
multilingual-e5-large | 79.44 |
Retrieval
MIRACL
Model | R@1 (%) ↑ | MRR@10 (%) ↑ | nDCG@10 (%) ↑ |
---|---|---|---|
SimCSE-IndoBERT Base | 36.04 | 48.25 | 39.70 |
ConGen-IndoBERT Lite Base | 46.04 | 59.06 | 51.01 |
ConGen-IndoBERT Base | 45.93 | 58.58 | 49.95 |
ConGen-SimCSE-IndoBERT Base | 45.83 | 58.27 | 49.91 |
ConGen-Indo-e5 Small | 55.00 | 66.74 | 58.95 |
SCT-IndoBERT Base | 40.41 | 47.29 | 40.68 |
all-IndoBERT Base | 65.52 | 75.92 | 70.13 |
all-IndoBERT Base-v2 | 67.18 | 76.59 | 70.16 |
all-Indo-e5 Small-v2 | 68.33 | 78.33 | 73.04 |
all-Indo-e5 Small-v3 | 68.12 | 78.22 | 73.09 |
distiluse-base-multilingual-cased-v2 | 41.35 | 54.93 | 48.79 |
paraphrase-multilingual-mpnet-base-v2 | 52.81 | 65.07 | 57.97 |
multilingual-e5-small | 70.20 | 79.61 | 74.80 |
multilingual-e5-base | 70.00 | 79.50 | 75.16 |
multilingual-e5-large | 70.83 | 80.58 | 76.16 |
TyDiQA
Model | R@1 (%) ↑ | MRR@10 (%) ↑ | nDCG@10 (%) ↑ |
---|---|---|---|
SimCSE-IndoBERT Base | 61.94 | 69.89 | 73.52 |
ConGen-IndoBERT Lite Base | 75.22 | 81.55 | 84.13 |
ConGen-IndoBERT Base | 73.09 | 80.32 | 83.29 |
ConGen-SimCSE-IndoBERT Base | 72.38 | 79.37 | 82.51 |
ConGen-Indo-e5 Small | 84.60 | 89.30 | 91.27 |
SCT-IndoBERT Base | 76.81 | 83.16 | 85.87 |
all-IndoBERT Base | 88.14 | 91.47 | 92.91 |
all-IndoBERT Base-v2 | 87.61 | 90.91 | 92.31 |
all-Indo-e5 Small-v2 | 93.27 | 95.63 | 96.46 |
all-Indo-e5 Small-v3 | 93.27 | 95.72 | 96.58 |
distiluse-base-multilingual-cased-v2 | 70.44 | 77.94 | 81.56 |
paraphrase-multilingual-mpnet-base-v2 | 81.41 | 87.05 | 89.44 |
multilingual-e5-small | 91.50 | 94.34 | 95.39 |
multilingual-e5-base | 93.45 | 95.88 | 96.69 |
multilingual-e5-large | 94.69 | 96.71 | 97.44 |
Classification
MTEB - Massive Intent Classification (id)
Model | Accuracy (%) ↑ | F1 Macro (%) ↑ |
---|---|---|
SimCSE-IndoBERT Base | 59.71 | 57.70 |
ConGen-IndoBERT Lite Base | 62.41 | 60.94 |
ConGen-IndoBERT Base | 61.14 | 60.02 |
ConGen-SimCSE-IndoBERT Base | 60.93 | 59.50 |
ConGen-Indo-e5 Small | 62.92 | 60.18 |
SCT-IndoBERT Base | 55.66 | 54.48 |
all-IndoBERT Base | 58.40 | 57.21 |
all-IndoBERT Base-v2 | 58.31 | 57.11 |
all-Indo-e5 Small-v2 | 61.51 | 59.24 |
all-Indo-e5 Small-v3 | 61.63 | 59.29 |
distiluse-base-multilingual-cased-v2 | 55.99 | 52.44 |
paraphrase-multilingual-mpnet-base-v2 | 65.43 | 63.55 |
multilingual-e5-small | 64.16 | 61.33 |
multilingual-e5-base | 66.63 | 63.88 |
multilingual-e5-large | 70.04 | 67.66 |
MTEB - Massive Scenario Classification (id)
Model | Accuracy (%) ↑ | F1 Macro (%) ↑ |
---|---|---|
SimCSE-IndoBERT Base | 66.14 | 65.56 |
ConGen-IndoBERT Lite Base | 67.25 | 66.53 |
ConGen-IndoBERT Base | 67.72 | 67.32 |
ConGen-SimCSE-IndoBERT Base | 67.12 | 66.64 |
ConGen-Indo-e5 Small | 66.92 | 66.29 |
SCT-IndoBERT Base | 61.89 | 60.97 |
all-IndoBERT Base | 66.37 | 66.31 |
all-IndoBERT Base-v2 | 66.02 | 65.97 |
all-Indo-e5 Small-v2 | 67.02 | 66.86 |
all-Indo-e5 Small-v3 | 67.27 | 67.13 |
distiluse-base-multilingual-cased-v2 | 65.25 | 63.45 |
paraphrase-multilingual-mpnet-base-v2 | 70.72 | 70.58 |
multilingual-e5-small | 67.92 | 67.23 |
multilingual-e5-base | 70.70 | 70.26 |
multilingual-e5-large | 74.11 | 73.82 |
IndoNLU - Emotion Classification (EmoT)
Model | Accuracy (%) ↑ | F1 Macro (%) ↑ |
---|---|---|
SimCSE-IndoBERT Base | 55.45 | 55.78 |
ConGen-IndoBERT Lite Base | 58.18 | 58.84 |
ConGen-IndoBERT Base | 57.04 | 57.06 |
ConGen-SimCSE-IndoBERT Base | 59.54 | 60.37 |
ConGen-Indo-e5 Small | 60.00 | 60.52 |
SCT-IndoBERT Base | 61.13 | 61.70 |
all-IndoBERT Base | 57.27 | 57.47 |
all-IndoBERT Base-v2 | 58.86 | 59.31 |
all-Indo-e5 Small-v2 | 58.18 | 57.99 |
all-Indo-e5 Small-v3 | 56.81 | 56.46 |
distiluse-base-multilingual-cased-v2 | 63.63 | 64.13 |
paraphrase-multilingual-mpnet-base-v2 | 63.18 | 63.78 |
multilingual-e5-small | 64.54 | 65.04 |
multilingual-e5-base | 68.63 | 69.07 |
multilingual-e5-large | 74.77 | 74.66 |
IndoNLU - Sentiment Analysis (SmSA)
Model | Accuracy (%) ↑ | F1 Macro (%) ↑ |
---|---|---|
SimCSE-IndoBERT Base | 85.6 | 81.50 |
ConGen-IndoBERT Lite Base | 81.2 | 75.59 |
ConGen-IndoBERT Base | 85.4 | 82.12 |
ConGen-SimCSE-IndoBERT Base | 83.0 | 78.74 |
ConGen-Indo-e5 Small | 84.2 | 80.21 |
SCT-IndoBERT Base | 82.0 | 76.92 |
all-IndoBERT Base | 84.4 | 79.79 |
all-IndoBERT Base-v2 | 83.4 | 79.04 |
all-Indo-e5 Small-v2 | 82.0 | 78.15 |
all-Indo-e5 Small-v3 | 82.6 | 78.98 |
distiluse-base-multilingual-cased-v2 | 78.8 | 73.64 |
paraphrase-multilingual-mpnet-base-v2 | 89.6 | 86.56 |
multilingual-e5-small | 83.6 | 79.51 |
multilingual-e5-base | 89.4 | 86.22 |
multilingual-e5-large | 90.0 | 86.50 |
Pair Classification
IndoNLI
Model | test_lay AP (%) ↑ |
test_expert AP (%) ↑ |
---|---|---|
SimCSE-IndoBERT Base | 56.06 | 50.72 |
ConGen-IndoBERT Lite Base | 69.44 | 53.74 |
ConGen-IndoBERT Base | 71.14 | 56.35 |
ConGen-SimCSE-IndoBERT Base | 70.80 | 56.59 |
ConGen-Indo-e5 Small | 70.51 | 55.67 |
SCT-IndoBERT Base | 59.82 | 53.41 |
all-IndoBERT Base | 72.01 | 56.79 |
all-IndoBERT Base-v2 | 71.36 | 56.83 |
all-Indo-e5 Small-v2 | 76.29 | 57.05 |
all-Indo-e5 Small-v3 | 75.21 | 56.62 |
distiluse-base-multilingual-cased-v2 | 58.48 | 50.50 |
paraphrase-multilingual-mpnet-base-v2 | 74.87 | 57.96 |
multilingual-e5-small | 63.97 | 51.85 |
multilingual-e5-base | 60.25 | 50.91 |
multilingual-e5-large | 61.39 | 51.62 |
Credits
Indonesian Sentence Embeddings is developed with love by: