import gradio as gr import spaces from transformers import pipeline import torch DESCRIPTION=""" This is the space for the Language Modeling Group at TABILAB in Computer Engineering of Bogazici University. We released the first version of our Turkish language model TURNA. This model is based on an encoder-decoder T5 architecture with 1.1B parameters. For more details, please refer to our paper. """ sentiment_example = [["Bu üründen çok memnun kaldım."]] long_text = [["Eyfel Kulesi (Fransızca: La tour Eiffel [la tuʀ ɛˈfɛl]), Paris'teki demir kule. Kule, aynı zamanda tüm dünyada Fransa'nın sembolü halini almıştır. İsmini, inşa ettiren Fransız inşaat mühendisi Gustave Eiffel'den alır.[1] En büyük turizm cazibelerinden biri olan Eyfel Kulesi, yılda 6 milyon turist çeker. 2002 yılında toplam ziyaretçi sayısı 200 milyona ulaşmıştır."]] ner_example = [["Benim adım Turna."]] t2t_example = [["Paraphrase: Bu üründen çok memnun kaldım."]] nli_example = [["Bunu çok beğendim. Bunu çok sevdim."]] #ttc = pipeline(model="boun-tabi-LMG/turna_classification_ttc4900", device=0) # examples =long_text, title="Text Categorization") #product_reviews = pipeline(model="boun-tabi-LMG/turna_classification_tr_product_reviews", device=0) #title_gen = pipeline(model="boun-tabi-LMG/turna_title_generation_mlsum", device=0) @spaces.GPU def nli(input, model_choice="turna_nli_nli_tr"): nli_model = pipeline(model="boun-tabi-LMG/turna_nli_nli_tr", device=0) stsb_model = pipeline(model="boun-tabi-LMG/turna_semantic_similarity_stsb_tr", device=0) if model_choice=="turna_nli_nli_tr": return nli_model(input)[0]["generated_text"] else: return stsb_model(input)[0]["generated_text"] @spaces.GPU def nli(input, model_choice="turna_nli_nli_tr"): nli_model = pipeline(model="boun-tabi-LMG/turna_nli_nli_tr", device=0) stsb_model = pipeline(model="boun-tabi-LMG/turna_semantic_similarity_stsb_tr", device=0) if model_choice=="turna_nli_nli_tr": return nli_model(input)[0]["generated_text"] else: return stsb_model(input)[0]["generated_text"] @spaces.GPU def sentiment_analysis(input, model_choice="turna_classification_17bintweet_sentiment"): product_reviews = pipeline(model="boun-tabi-LMG/turna_classification_tr_product_reviews", device=0) sentiment_model = pipeline(model="boun-tabi-LMG/turna_classification_17bintweet_sentiment", device=0) if model_choice=="turna_classification_17bintweet_sentiment": return sentiment_model(input)[0]["generated_text"] else: return product_reviews(input)[0]["generated_text"] @spaces.GPU def t2t(input): return t2t_gen_model(input) @spaces.GPU def pos(input, model_choice="turna_pos_imst"): if model_choice=="turna_pos_imst": pos_imst = pipeline(model="boun-tabi-LMG/turna_pos_imst", device=0) return pos_imst(input)[0]["generated_text"] else: pos_boun = pipeline(model="boun-tabi-LMG/turna_pos_boun", device=0) return pos_boun(input)[0]["generated_text"] @spaces.GPU def ner(input, model_choice="turna_ner_wikiann"): if model_choice=="turna_ner_wikiann": ner_wikiann = pipeline(model="boun-tabi-LMG/turna_ner_wikiann", device=0) return ner_wikiann(input)[0]["generated_text"] else: ner_model = pipeline(model="boun-tabi-LMG/turna_ner_milliyet", device=0) return ner_model(input)[0]["generated_text"] @spaces.GPU def paraphrase(input, model_choice="turna_paraphrasing_tatoeba"): paraphrasing = pipeline(model="boun-tabi-LMG/turna_paraphrasing_tatoeba", device=0) paraphrasing_sub = pipeline(model="boun-tabi-LMG/turna_paraphrasing_opensubtitles", device=0) if model_choice=="turna_paraphrasing_tatoeba": return paraphrasing(input)[0]["generated_text"] else: return paraphrasing_sub(input)[0]["generated_text"] @spaces.GPU def summarize(input, model_choice="turna_summarization_tr_news"): summarization_model = pipeline(model="boun-tabi-LMG/turna_summarization_mlsum", device=0) news_sum = pipeline(model="boun-tabi-LMG/turna_summarization_tr_news", device=0) if model_choice=="turna_summarization_tr_news": return news_sum(input)[0]["generated_text"] else: return summarization_model(input)[0]["generated_text"] with gr.Blocks(theme="soft") as demo: gr.Markdown("# TURNA 🐦") gr.Markdown(DESCRIPTION) with gr.Tab("POS"): gr.Markdown("TURNA fine-tuned on part-of-speech-tagging. Enter text to parse parts of speech and pick the model.") with gr.Column(): with gr.Row(): pos_choice = gr.Radio(choices = ["turna_pos_imst", "turna_pos_boun"], label ="Model") pos_input = gr.Textbox(label="POS Input") pos_output = gr.Textbox(label="POS Output") pos_submit = gr.Button() pos_submit.click(pos, inputs=[pos_input, pos_choice], outputs=pos_output) pos_examples = gr.Examples(examples = ner_example, inputs = [pos_input, pos_choice], outputs=pos_output, fn=pos) with gr.Tab("NER"): gr.Markdown("TURNA fine-tuned on named entity recognition. Enter text to parse named entities and pick the model.") with gr.Column(): with gr.Row(): ner_choice = gr.Radio(choices = ["turna_ner_wikiann", "turna_ner_milliyet"], label ="Model") ner_input = gr.Textbox(label="NER Input") ner_submit = gr.Button() ner_output = gr.Textbox(label="NER Output") ner_submit.click(ner, inputs=[ner_input, ner_choice], outputs=ner_output) ner_examples = gr.Examples(examples = ner_example, inputs = [ner_input, ner_choice], outputs=ner_output, fn=ner) with gr.Tab("Paraphrase"): gr.Markdown("TURNA fine-tuned on paraphrasing. Enter text to paraphrase and pick the model.") with gr.Column(): with gr.Row(): paraphrasing_choice = gr.Radio(choices = ["turna_paraphrasing_tatoeba", "turna_paraphrasing_opensubtitles"], label ="Model") paraphrasing_input = gr.Textbox(label = "Paraphrasing Input") paraphrasing_submit = gr.Button() paraphrasing_output = gr.Text(label="Paraphrasing Output") paraphrasing_submit.click(paraphrase, inputs=[paraphrasing_input, paraphrasing_choice], outputs=paraphrasing_output) paraphrase_examples = gr.Examples(examples = long_text, inputs = [paraphrasing_input, paraphrasing_choice], outputs=paraphrasing_output, fn=paraphrase) with gr.Tab("Summarization"): gr.Markdown("TURNA fine-tuned on summarization. Enter text to summarize and pick the model.") with gr.Column(): with gr.Row(): sum_choice = gr.Radio(choices = ["turna_summarization_mlsum", "turna_summarization_tr_news"], label ="Model") sum_input = gr.Textbox(label = "Summarization Input") sum_submit = gr.Button() sum_output = gr.Textbox(label = "Summarization Output") sum_submit.click(summarize, inputs=[sum_input, sum_choice], outputs=sum_output) sum_examples = gr.Examples(examples = long_text, inputs = [sum_input, sum_choice], outputs=sum_output, fn=summarize) demo.launch()