File size: 18,574 Bytes
6e74145
 
 
 
a40a80d
 
 
 
 
 
 
 
83a1143
 
 
 
6e74145
 
 
 
 
 
 
a40a80d
 
 
6e74145
 
 
 
a40a80d
6e74145
 
 
a40a80d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6e74145
a40a80d
 
 
6e74145
a40a80d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
83a1143
a40a80d
 
 
 
 
 
 
 
1ccde3b
a40a80d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d4e5279
6e74145
d4e5279
6e74145
 
 
 
 
 
 
 
a40a80d
6e74145
 
 
a40a80d
 
 
 
 
6e74145
 
 
 
 
 
 
 
 
 
6c8898d
d4e5279
6e74145
 
 
a40a80d
 
 
 
 
 
 
 
d4e5279
 
a40a80d
 
d4e5279
1ccde3b
 
 
d4e5279
 
 
 
 
 
 
 
 
 
 
 
6e74145
a40a80d
 
6e74145
a40a80d
 
 
 
 
 
 
 
d4e5279
 
a40a80d
6e74145
a40a80d
 
6e74145
 
 
 
 
 
 
 
a40a80d
6e74145
a40a80d
6e74145
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
83a1143
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1ccde3b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
83a1143
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1ccde3b
83a1143
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1ccde3b
 
83a1143
 
 
 
 
 
 
 
 
 
1ccde3b
83a1143
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6e74145
a40a80d
d4e5279
83a1143
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
import gradio as gr
import torch
import itertools
import pandas as pd
import spaces
import random
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModel
from sklearn.metrics import pairwise_distances
from collections import Counter
from itertools import chain
from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction
import math
import markdown
from .text import doctree_from_url, get_selectors_for_class, split_by_heading, DocTree
from .optimization import ngrams, count_ngrams, self_bleu, dist_n, perplexity, js_divergence


model_name = 'philipp-zettl/t5-small-long-qa'
qa_model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
model_name = 'philipp-zettl/t5-small-qg'
qg_model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained('google/flan-t5-small')

embedding_model = AutoModel.from_pretrained('sentence-transformers/paraphrase-MiniLM-L6-v2')
embedding_tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/paraphrase-MiniLM-L6-v2')

# Move only the student model to GPU if available
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
qa_model = qa_model.to(device)
qg_model = qg_model.to(device)
embedding_model = embedding_model.to(device)

max_questions = 1
max_answers = 1
max_elem_value = 100


def embedding_similarity(inputs, outputs):
    global embedding_model, embedding_tokenizer, device
    def embed(texts):
        inputs = embedding_tokenizer(texts, return_tensors='pt', padding=True, truncation=True).to(device)
        with torch.no_grad():
            outputs = embedding_model(**inputs)
        return outputs.last_hidden_state.mean(dim=1).cpu().numpy()

    input_embeddings = embed(inputs)
    output_embeddings = embed(outputs)

    similarities = pairwise_distances(input_embeddings, output_embeddings, metric='cosine')
    return sum(similarities) / len(similarities)


def evaluate_model(num_beams, num_beam_groups, model, tokenizer, eval_data, max_length=85):
    generated_outputs = []

    for input_text in eval_data:
        input_ids = tokenizer.encode(input_text, return_tensors="pt").to(device)
        outputs = model.generate(
            input_ids, 
            num_beams=num_beams, 
            num_beam_groups=num_beam_groups, 
            diversity_penalty=1.0,
            max_new_tokens=max_length,
        )
        decoded_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
        generated_outputs.append(decoded_text.split())

    # Self-BLEU for diversity
    diversity_score = self_bleu(generated_outputs)

    # Dist-1 and Dist-2 for diversity
    dist1 = dist_n(generated_outputs, 1)
    dist2 = dist_n(generated_outputs, 2)

    # Perplexity for fluency and relevance
    fluency_score = perplexity(model, tokenizer, [" ".join(output) for output in generated_outputs])

    # Embedding similarity for contextual relevance
    contextual_score = embedding_similarity(eval_data, [" ".join(output) for output in generated_outputs])

    # Jensen-Shannon Divergence for distribution similarity
    generated_ngrams = count_ngrams(list(chain(*generated_outputs)), 4)
    reference_ngrams = count_ngrams(list(chain(*[tokenizer.tokenize(text) for text in eval_data])), 4)
    all_ngrams = set(generated_ngrams.keys()).union(set(reference_ngrams.keys()))
    p = [generated_ngrams[ngram] for ngram in all_ngrams]
    q = [reference_ngrams[ngram] for ngram in all_ngrams]
    jsd_score = js_divergence(p, q)

    return {
        "diversity_score": diversity_score,
        "dist1": dist1,
        "dist2": dist2,
        "fluency_score": fluency_score,
        "contextual_score": contextual_score,
        "jsd_score": jsd_score
}


def find_best_parameters(eval_data, model, tokenizer, max_length=85):

    # Parameter ranges
    parameter_map = {
        2: [2],
        4: [2],
        6: [2], # 6x3 == 4x2
        8: [2], # 8x4 == 6x3 == 4x2
        9: [3],
        10: [2], # 10x5 == 8x4 == 6x3 == 4x2
    }

    # Find the best parameters
    best_score = -float('inf')
    best_params = None

    for num_beams in parameter_map.keys():
        for num_beam_groups in parameter_map[num_beams]:
            if num_beam_groups > num_beams:
                continue  # num_beam_groups should not be greater than num_beams

            scores = evaluate_model(num_beams, num_beam_groups, model, tokenizer, eval_data, max_length=max_length)
            # Combine scores to determine the best parameters
            combined_score = (scores['dist1'] + scores['dist2'] - scores['fluency_score'] + scores['contextual_score'] - scores['jsd_score']).mean()
            print(f"num_beams={num_beams}, num_beam_groups={num_beam_groups}, avg combined score={combined_score}")
            
            if combined_score > best_score:
                best_score = combined_score
                best_params = (num_beams, num_beam_groups)

    print(f"Best parameters: num_beams={best_params[0]}, num_beam_groups={best_params[1]} with combined score={best_score}")
    return best_params


def run_model(inputs, tokenizer, model, num_beams=2, num_beam_groups=2, temperature=0.5, num_return_sequences=1, max_length=85, seed=42069):
    all_outputs = []
    torch.manual_seed(seed)
    for input_text in inputs:
        model_inputs = tokenizer([input_text], max_length=512, padding=True, truncation=True)
        input_ids = torch.tensor(model_inputs['input_ids']).to(device)
        for sample in input_ids:
            sample_outputs = []
            with torch.no_grad():
                sample_output = model.generate(
                    input_ids[:1],
                    max_length=max_length,
                    num_return_sequences=num_return_sequences,
                    low_memory=True,
                    use_cache=True,
                    # Diverse Beam search decoding
                    num_beams=max(2, num_return_sequences),
                    num_beam_groups=max(2, num_return_sequences),
                    diversity_penalty=temperature,

                )
                for i, sample_output in enumerate(sample_output):
                    sample_output = sample_output.unsqueeze(0)
                    sample_output = tokenizer.decode(sample_output[0], skip_special_tokens=True)
                    sample_outputs.append(sample_output)

            all_outputs.append(sample_outputs)
    return all_outputs


@spaces.GPU
def gen(content, temperature_qg=0.5, temperature_qa=0.75, num_return_sequences_qg=1, num_return_sequences_qa=1, max_length=85, seed=42069, optimize_questions=False):
    inputs = [
        f'context: {content}'
    ]
    question = run_model(
        inputs, 
        tokenizer, 
        qg_model, 
        num_beams=num_return_sequences_qg, 
        num_beam_groups=num_return_sequences_qg, 
        temperature=temperature_qg, 
        num_return_sequences=num_return_sequences_qg, 
        max_length=max_length,
        seed=seed
    )

    if optimize_questions:
        q_params = find_best_parameters(
            list(chain.from_iterable(question)), qg_model, tokenizer, max_length=max_length
        )

        question = run_model(
            inputs, 
            tokenizer, 
            qg_model, 
            num_beams=q_params[0],
            num_beam_groups=q_params[1],
            temperature=temperature_qg, 
            num_return_sequences=num_return_sequences_qg, 
            max_length=max_length,
            seed=seed
        )

    inputs = list(chain.from_iterable([
        [f'question: {q} context: {content}' for q in q_set] for q_set in question
    ]))
    answer = run_model(
        inputs,
        tokenizer, 
        qa_model, 
        num_beams=num_return_sequences_qa,
        num_beam_groups=num_return_sequences_qa,
        temperature=temperature_qa,
        num_return_sequences=num_return_sequences_qa,
        max_length=max_length,
        seed=seed
    )

    questions = list(chain.from_iterable(question))
    answers = list(chain.from_iterable(answer))

    results = []
    for idx, ans in enumerate(answers):
        results.append({'question': questions[idx % num_return_sequences_qg], 'answer': ans})
    return results


def variable_outputs(k, max_elems=10):
    global max_elem_value
    k = int(k)
    return [gr.Text(visible=True)] * k + [gr.Text(visible=False)] * (max(max_elems, max_elem_value)- k)


def set_outputs(content, max_elems=10):
    c = eval(content)
    print('received content: ', c)
    return [gr.Text(value=t, visible=True) for t in c] + [gr.Text(visible=False)] * (max(max_elems, 10) - len(c))


def create_file_download(qnas):
    with open('qnas.tsv', 'w') as f:
        for idx, qna in qnas.iterrows():
            f.write(qna['Question'] + '\t' + qna['Answer'])
            if idx < len(qnas) - 1:
                f.write('\n')
    return 'qnas.tsv'


def main():
    with gr.Tab(label='QA Generator'):
        with gr.Tab(label='Explanation'):
            gr.Markdown(
                '''
                # QA Generator
                This tab allows you to generate questions and answers from a given piece of text content.

                ## How to use
                1. Enter the text content you want to generate questions and answers from.
                2. Adjust the diversity penalty for question generation and answer generation.
                3. Set the maximum length of the generated questions and answers.
                4. Choose the number of questions and answers you want to generate.
                5. Click on the "Generate" button.

                The next section will give you insights into the generated questions and answers.

                If you're satisfied with the generated questions and answers, you can download them as a TSV file.
                '''
            )
            with gr.Accordion(label='Optimization', open=False):
                gr.Markdown("""
                For optimization of the question generation we apply the following combined score:

                $$\\text{combined} = \\text{dist1} + \\text{dist2} - \\text{fluency} + \\text{contextual} - \\text{jsd}$$

                Here's a brief explanation of each component:

                1. **dist1 and dist2**: These represent the diversity of the generated outputs. dist1 measures the ratio of unique unigrams to total unigrams, and dist2 measures the ratio of unique bigrams to total bigrams. <u>**Higher values indicate more diverse outputs.**</u>

                2. **fluency**: This is the perplexity of the generated outputs, which measures how well the outputs match the language model's expectations. <u>**Lower values indicate better fluency.**</u>

                3. **contextual**: This measures the similarity between the input and generated outputs using embedding similarity. <u>**Higher values indicate better contextual relevance.**</u>

                4. **jsd**: This is the Jensen-Shannon Divergence between the n-gram distributions of the generated outputs and the reference data. <u>**Lower values indicate greater similarity between distributions.**</u>
                """, latex_delimiters=[{'display': False, 'left': '$$', 'right': '$$'}])
        with gr.Tab(label='Generate QA'):
            with gr.Row(equal_height=True):
                with gr.Group("Content"):
                    content = gr.Textbox(label='Content', lines=15, placeholder='Enter text here', max_lines=10_000)
                with gr.Group("Settings"):
                    temperature_qg = gr.Slider(label='Diversity Penalty QG', value=0.2, minimum=0, maximum=1, step=0.01)
                    temperature_qa = gr.Slider(label='Diversity Penalty QA', value=0.5, minimum=0, maximum=1, step=0.01)
                    max_length = gr.Number(label='Max Length', value=85, minimum=1, step=1, maximum=512)
                    num_return_sequences_qg = gr.Number(label='Number Questions', value=max_questions, minimum=1, step=1, maximum=max(max_questions, max_elem_value))
                    num_return_sequences_qa = gr.Number(label="Number Answers", value=max_answers, minimum=1, step=1, maximum=max(max_questions, max_elem_value))
                    seed = gr.Number(label="seed", value=42069)
                    optimize_questions = gr.Checkbox(label="Optimize questions?", value=False)

            with gr.Row():
                gen_btn = gr.Button("Generate")

            @gr.render(
                inputs=[
                    content, temperature_qg, temperature_qa, num_return_sequences_qg, num_return_sequences_qa,
                    max_length, seed, optimize_questions
                ],
                triggers=[gen_btn.click]
            )
            def render_results(content, temperature_qg, temperature_qa, num_return_sequences_qg, num_return_sequences_qa, max_length, seed, optimize_questions):
                if not content.strip():
                    raise gr.Error('Please enter some content to generate questions and answers.')
                qnas = gen(
                    content, temperature_qg, temperature_qa, num_return_sequences_qg, num_return_sequences_qa,
                    max_length, seed, optimize_questions
                )
                df = gr.Dataframe(
                    value=[u.values() for u in qnas],
                    headers=['Question', 'Answer'],
                    col_count=2,
                    wrap=True
                )
                pd_df = pd.DataFrame([u.values() for u in qnas], columns=['Question', 'Answer'])

                download = gr.DownloadButton(label='Download (without headers)', value=create_file_download(pd_df))

            content.change(lambda x: x.strip(), content)


def new_main():
    with gr.Tab('Content extraction from URL'):
        with gr.Tab(label='Explanation'):
            gr.Markdown(
                '''
                # Content extraction from URL
                This tab allows you to extract content from a URL and chunk it into sections.

                ## How to use
                1. Enter the URL of the webpage you want to extract content from.
                2. Select the element class and class name of the HTML element you want to extract content from.
                3. Click on the "Extract content" button.

                The next section will give you insights into the extracted content.

                This was done to give you the possibility to look at the extracted content, as well as manipulate it further.

                Once you extract the content, you can choose the depth level to chunk the content into sections.
                1. Enter the depth level you want to chunk the content into. **Note: <u>This is based on the HTML structure of the webpage, we're utilizing heading tags for this purpose</u>**
                2. Click on the "Chunk content" button.
                '''
            )
        with gr.Tab(label='Extract content'):
            url = gr.Textbox(label='URL', placeholder='Enter URL here', lines=1, max_lines=1)
            elem_class = gr.Dropdown(label='CSS element class', choices=['div', 'p', 'span'], value='div')
            class_name = gr.Dropdown(label='CSS class name', choices=[], allow_custom_value=True)

            extract_btn = gr.Button('Extract content')

            with gr.Group():
                content_state = gr.State(None)
                final_content = gr.Textbox(value='', show_copy_button=True, label='Final content', interactive=True)
                with gr.Accordion('Reveal original input', open=False):
                    og_content = gr.Textbox(value='', label='OG HTML content')

            with gr.Group(visible=False) as step_2_group:
                depth_level = gr.Number(label='Depth level', value=1, minimum=0, step=1, maximum=6)
                continue_btn = gr.Button('Chunk content')

            def render_results(url, elem_class_, class_name_):
                if not url.strip():
                    raise gr.Error('Please enter a URL to extract content.')
                content = doctree_from_url(url, elem_class_, class_name_)
                return [
                    content,
                    content.content,
                    content.as_markdown(content.merge_sections(content.get_sections(0))),
                    gr.Group(visible=True)
                ]

            def get_class_options(url, elem_class):
                if not url.strip():
                    raise gr.Error('Please enter a URL to extract content.')

                return gr.Dropdown(label='CSS class name', choices=list(set(get_selectors_for_class(url, elem_class))))

            def update_content_state_on_final_change(final_content):
                html_content = markdown.markdown(final_content)
                return DocTree(split_by_heading(html_content, 1))

            @gr.render(inputs=[content_state, depth_level], triggers=[continue_btn.click])
            def select_content(content, depth_level):
                if not content:
                    raise gr.Error('Please extract content first.')

                sections = content.get_sections_by_depth(depth_level)
                print(f'Found {len(sections)} sections')
                ds = []
                for idx, section in enumerate(sections):
                    ds.append([idx, content.as_markdown(content.merge_sections(section))])
                gr.Dataframe(value=ds, headers=['Section #', 'Content'], interactive=True, wrap=True)

            url.change(
                get_class_options,
                inputs=[url, elem_class],
                outputs=[class_name]
            )

            extract_btn.click(
                render_results,
                inputs=[
                    url, elem_class, class_name,
                ],
                outputs=[
                    content_state, og_content, final_content, step_2_group
                ]
            )
            final_content.change(update_content_state_on_final_change, inputs=[final_content], outputs=[content_state])


with gr.Blocks() as demo:
    gr.Markdown(
        '''
        # QA-Generator
        A tool to build FAQs or QnAs from a given piece of text content.

        ## How to use
        We provide you two major functionalities:
        1. **Content extraction from URL**: Extract content from a URL and chunk it into sections.
        2. **QA Generator**: Generate questions and answers from a given text content.

        Select the tab you want to use and follow the instructions.
        '''
    )
    new_main()
    main()


demo.queue()
demo.launch()