Polyglot or Not? Measuring Multilingual Encyclopedic Knowledge in Foundation Models

This is the repository for the following paper: Polyglot or Not? Measuring Multilingual Encyclopedic Knowledge in Foundation Models, which was published in EMNLP 2023 (Main). It contains several research artifacts, including:

1. The code for running the fact-completion test
2. Our dataset of factual associations translated into 20 languages
3. A demo of contrastive knowledge assessment

Method

Given a factual association such as The capital of France is *Paris***, we determine whether a model adequately "knows" the correct completion with the following test:

• Step 1: prompt the model to predict the likelihood of the token Paris following The Capital of France is
• Step 2: prompt the model to predict the average likelihood of a set of false, counterfactual tokens following the same stem.

If the value from Step 1 is greater than the value from Step 2 we conclude that the model adequately recalls that fact. Formally, this is an application of the Contrastive Knowledge Assessment proposed in [1].

Models Evaluated

We evaluate 5 foundation models of interest in a multilingual setting, like Llama [2]. We perform this assessment with 303k fact-completions spanning 20 languages (results).

In addition to our multilingual assessment, we also scored a diverse set of ~30 models (like Mistral, Llama-2, and Falcon) on the English-only subset of our dataset, which comprises 26.3k fact-completions.

While we would have liked to test close-sourced models, such as OpenAI's GPT-4, such models do not provide vocabulary-wide probabilities at inference. These models are thus incompatible at present with our contrastive knowledge assessment test. As such, our study demonstrates the need for all LLMs - open and closed - to produce vocabulary-wide probabilities for more robust evaluations.

Data Release

We present 303k unique fact-completions in Polyglot-or-Not/Fact-Completion, which are in the form of {stem, fact, counterfact} triples. See the dataset viewer for a closer look.

• 20 Latin/Cyrillic script languages are included. The ISO 639-1 language codes are: bg, ca, cs, da, de, en, es, fr, hr, hu, it, nl, pl, pt, ro, ru, sl, sr, sv, and uk.

The factual associations were originally sourced from English-language Wikidata curated in the T-REx dataset [3] as utilized in factual association research such as [1] and [4]. We used the Google Translate API alongside bespoke wrapper code to programmatically generate the non-English cuts.

Test Results

### Multilingual leaderboard

| model | accuracy (%) | params | n tokens |------------------|:--------------:|:--------------:|:--------------:| | llama-33b | 79.31 (+/- 0.74) | 32.5B | 1.4T | | m-bert | 62.00 (+/- 0.87) | 110M | - | | bloom-7b1 | 57.70 (+/- 0.88) | 7.1B | 341B | | xlm-roberta | 56.03 (+/- 0.90) | 355M | 295B | | mt5-xl | 52.51 (+/- 0.91) | 3.7B | - | | Random Baseline | 50 | - | - |

Table 1: Multilingual test leaderboard. Here, accuracy refers to the average performance of each model across 20 distinct languages. The uncertainty estimates represent averaged 95% confidence intervals computed from 10000 bootstrap iterations per language. Params and n tokens record each models number of parameters and number of dataset tokens, respectively (when such data is available). These results reveal that models struggle to recall facts in a multilingual setting, as compared to their English-only performance (Table 2). For instance, on average, Llama-33B's accuracy decreased by approximately 11% from English to non-English languages.

### English-only leaderboard

| model | accuracy (%) | params | n tokens |------------------|:--------------:|:--------------:|:--------------:| | falcon-180b | 91.53 (+/- 0.34) | 180B | 3.5T | | llama-2-70b | 90.86 (+/- 0.35) | 70B | 2T | | llama-65b | 89.56 (+/- 0.37) | 65.2B | 1.4T | | llama-33b | 89.40 (+/- 0.38) | 32.5B | 1.4T | | llama-2-13b | 87.51 (+/- 0.40) | 13B | 2T | | falcon-40b | 87.01 (+/- 0.41) | 40B | 1T | | mistral-7b-v0.1 | 86.88 (+/- 0.41) | 7.3B | - | | llama-13b | 86.66 (+/- 0.42) | 12.5B | 1T | | llama-2-7b | 86.22 (+/- 0.42) | 7B | 2T | | llama-7b | 85.53 (+/- 0.43) | 6.7B | 1T | | mpt-30b | 85.09 (+/- 0.43) | 30B | 1T | | redpajama-7b | 85.07 (+/- 0.44) | 7B | 800B | | mpt-7b | 83.39 (+/- 0.46) | 7B | 1T | | opt-30b | 82.09 (+/- 0.47) | 30B | 180B | | redpajama-3b | 82.09 (+/- 0.47) | 3B | 800B | | opt-13b | 81.94 (+/- 0.46) | 13B | 30B | | gpt-neox-20b | 81.50 (+/- 0.47) | 20B | 420B | | falcon-7b | 81.34 (+/- 0.47) | 7B | 1.5T | | gpt-j-6b | 81.14 (+/- 0.47) | 6B | 420B | | pythia-12b | 80.53 (+/- 0.48) | 12B | 420B | | t5-v1-xxl | 76.55 (+/- 0.52) | 13B | 34B | | bloom-7b1 | 76.16 (+/- 0.51) | 7.1B | 341B | | gpt2-xl | 73.76 (+/- 0.54) | 1.5B | - | | bert | 72.60 (+/- 0.54) | 110M | - | | m-bert | 71.80 (+/- 0.55) | 110M | - | | stablelm-7b | 68.85 (+/- 0.55) | 7B | 1.5T | | distilgpt2 | 64.23 (+/- 0.59) | 82M | - | | mt5-xxl | 61.58 (+/- 0.59) | 13B | - | | xlm-roberta | 61.55 (+/- 0.59) | 355M | 295B | | mt5-xl | 59.96 (+/- 0.59) | 3.7B | - | | Random Baseline | 50 | - | - |

Table 2: Monolingual test leaderboard. Accuracy represents performance on English-only data. The uncertainty estimates are 95% confidence intervals computed from 10000 bootstrap iterations. Params and n tokens record each models number of parameters and number of dataset tokens, respectively (when such data is available). Consistent with the trends in Table 1, Llamas of varying sizes emerge as the front-runners.

Llama-33B performance across languages

Figure 1: Llama-33B's test performance across languages. Accuracy denotes the model's performance assessed individually for each language. The Llama-33B model demonstrates higher proficiency with languages utilizing the Latin script as compared to those using the Cyrillic script (Ukrainian, Bulgarian, Russian, and Serbian). A chi-squared test substantiates a significant dependency of the model's test performance on the language script (2 = 3570.576, p < 0.001).

Authors

• Daniel Furman
• Tim Schott timschott@berkeley.edu
• Shreshta Bhat

Advisor

• David Bamman dbamman@berkeley.edu

Citation

Please cite this repository as follows if you use its data or code:

@misc{schott2023polyglot, doi = {10.48550/arXiv.2305.13675}, title={Polyglot or Not? Measuring Multilingual Encyclopedic Knowledge in Foundation Models}, author={Tim Schott and Daniel Furman and Shreshta Bhat}, year={2023}, eprint={2305.13675}, archivePrefix={arXiv}, primaryClass={cs.CL} }

Bibliography

[1] Calibrating Factual Knowledge in Pretrained Language Models. Dong, Qingxiu, Damai Dai, Yifan Song, Jingjing Xu, Zhifang Sui, and Lei Li. In Findings of the Association for Computational Linguistics: EMNLP 2022. arXiv:2210.03329 (2022).

[2] Llama: Open and Efficient Foundation Language Models. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothe Lacroix, Baptiste Rozire, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, Guillaume Lample. https://arxiv.org/abs/2302.13971v1 (2023).

• Llama weights were accessed with the approval of Meta AI and used in accordance with the License (see link for more details).
  

[3] T-REx: A Large Scale Alignment of Natural Language with Knowledge Base Triples. ElSahar, Hady, Pavlos Vougiouklis, Arslen Remaci, Christophe Gravier, Jonathon S. Hare, Frdrique Laforest and Elena Paslaru Bontas Simperl. International Conference on Language Resources and Evaluation. Link (2018).

[4] Mass Editing Memory in a Transformer. Meng, Kevin, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. arXiv preprint arXiv:2210.07229 (2022).