Anirban Bhowmick

Mehdi Ali, Michael Fromm, Klaudia Thellmann et al.

We present two multilingual LLMs, Teuken 7B-base and Teuken 7B-instruct, designed to embrace Europe's linguistic diversity by supporting all 24 official languages of the European Union. Trained on a dataset comprising around 60% non-English data and utilizing a custom multilingual tokenizer, our models address the limitations of existing LLMs that predominantly focus on English or a few high-resource languages. We detail the models' development principles, i.e., data composition, tokenizer optimization, and training methodologies. The models demonstrate strong performance across multilingual benchmarks, as evidenced by their performance on European versions of ARC, HellaSwag, and TruthfulQA.

Nicolo' Brandizzi, Hammam Abdelwahab, Anirban Bhowmick et al.

This paper presents a comprehensive overview of the data preparation pipeline developed for the OpenGPT-X project, a large-scale initiative aimed at creating open and high-performance multilingual large language models (LLMs). The project goal is to deliver models that cover all major European languages, with a particular focus on real-world applications within the European Union. We explain all data processing steps, starting with the data selection and requirement definition to the preparation of the final filtered data. We distinguish between curated data and web data, as each of these categories is handled by distinct pipelines, with curated data undergoing minimal filtering and web data requiring extensive filtering and deduplication. This distinction guided the development of specialized algorithmic solutions for both pipelines. In addition to describing the processing methodologies, we provide an in-depth analysis of the datasets, increasing transparency and alignment with European data regulations. Finally, we share key insights and challenges faced during the project, offering recommendations for future endeavors in large-scale multilingual data preparation for LLMs.

Niklas von Boguszewski, Sana Moin, Anirban Bhowmick et al.

In this research, we investigate techniques to detect hate speech in movies. We introduce a new dataset collected from the subtitles of six movies, where each utterance is annotated either as hate, offensive or normal. We apply transfer learning techniques of domain adaptation and fine-tuning on existing social media datasets, namely from Twitter and Fox News. We evaluate different representations, i.e., Bag of Words (BoW), Bi-directional Long short-term memory (Bi-LSTM), and Bidirectional Encoder Representations from Transformers (BERT) on 11k movie subtitles. The BERT model obtained the best macro-averaged F1-score of 77%. Hence, we show that transfer learning from the social media domain is efficacious in classifying hate and offensive speech in movies through subtitles.

Anirban Bhowmick, Astik Biswas

language identification (LID) is identifing a language in a given spoken utterance. Language segmentation is equally inportant as language identification where language boundaries can be spotted in a multi language utterance. In this paper, we have experimented with two schemes for language identification in Indian regional language context as very few works has been done. Singular value based feature embedding is used for both of the schemes. In first scheme, the singular value decomposition (SVD) is applied to the n-gram utterance matrix and in the second scheme, SVD is applied on the difference supervector matrix space. We have observed that in both the schemes, 55-65% singular value energy is sufficient to capture the language context. In n-gram based feature representation, we have seen that different skipgram models capture different language context. We have observed that for short test duration, supervector based feature representation is better but with a longer duration test signal, n-gram based feature performed better. We have also extended our work to explore language-based segmentation where we have seen that segmentation accuracy of four language group with ten language training model, scheme-1 has performed well but with same four language training model, scheme-2 outperformed scheme-1

Srikanth Raj Chetupalli, Anirban Bhowmick, Thippur V. Sreenivas

Diarization of audio recordings from ad-hoc mobile devices using spatial information is considered in this paper. A two-channel synchronous recording is assumed for each mobile device, which is used to compute directional statistics separately at each device in a frame-wise manner. The recordings across the mobile devices are asynchronous, but a coarse synchronization is performed by aligning the signals using acoustic events, or real-time clock. Direction statistics computed for all the devices, are then modeled jointly using a Dirichlet mixture model, and the posterior probability over the mixture components is used to derive the diarization information. Experiments on real life recordings using mobile phones show a diarization error rate of less than 14%.