precomputation (#19)

- Add precomputing acl data (44257a91d79656a35a3eac347f3d60c5c66a1ef9)
- Add precomputing acl data (02d7398f2f7c165b2d5110ff7580990b0a4f19f5)


Co-authored-by: Yixi Ding <dyxohjl666@users.noreply.huggingface.co>

app.py

@@ -5,12 +5,13 @@ from reference_string_parsing import *
 from controlled_summarization import *
 from dataset_extraction import *
 
+from controlled_summarization import recommended_kw
 import requests
 
 # Example Usage
-#url = "https://arxiv.org/pdf/2305.14996.pdf"
-#dest_folder = "./examples/"
-#download_pdf(url, dest_folder)
+# url = "https://arxiv.org/pdf/2305.14996.pdf"
+# dest_folder = "./examples/"
+# download_pdf(url, dest_folder)
 
 
 with gr.Blocks(css="#htext span {white-space: pre-line}") as demo:
@@ -31,17 +32,20 @@ with gr.Blocks(css="#htext span {white-space: pre-line}") as demo:
                             gr.Markdown("* Set the length of text used for summarization. Length 0 will exert no control over length.")
                             # ctrlsum_file_beams = gr.Number(label="Number of beams for beam search", value=1, precision=0)
                             # ctrlsum_file_sequences = gr.Number(label="Number of generated summaries", value=1, precision=0)
-                            ctrlsum_file_length = gr.Slider(0,300,step=50, label="Length")
-                            ctrlsum_file_keywords = gr.Textbox(label="Keywords",max_lines=1)
+                            ctrlsum_file_length = gr.Radio(label="Length", value=0, choices=[0, 50, 100, 200])
+                            kw = gr.Radio(visible=False)
+                            ctrlsum_file_keywords = gr.Textbox(label="Keywords", max_lines=1)
                         with gr.Row():
                             ctrlsum_file_btn = gr.Button("Generate")
                     ctrlsum_file_output = gr.Textbox(
                         elem_id="htext",
                         label="Summary",
                     )
-                ctrlsum_file_examples = gr.Examples(examples=[["examples/H01-1042_body.txt", 50, "automatic evaluation technique", "",""],["examples/H01-1042.pdf", 0, "automatic evaluation technique","",""]],
-                                                inputs=[ctrlsum_file, ctrlsum_file_length, ctrlsum_file_keywords, ctrlsum_str, ctrlsum_url])
-
+                ctrlsum_file_examples = gr.Examples(
+                    examples=[["examples/H01-1042_body.txt", 50, "automatic evaluation technique", "", ""],
+                              ["examples/H01-1042.pdf", 0, "automatic evaluation technique", "", ""]],
+                    inputs=[ctrlsum_file, ctrlsum_file_length, ctrlsum_file_keywords, ctrlsum_str, ctrlsum_url
+                            ])
 
 
 
@@ -51,13 +55,34 @@ with gr.Blocks(css="#htext span {white-space: pre-line}") as demo:
             outputs=[ctrlsum_file_output, ctrlsum_str, ctrlsum_file]
         )
         def clear():
-            return None,0,None, None
+            return None, 0, None, None, gr.Radio(visible=False)
+
+
+        def update_url(url):
+            if url in recommended_kw.keys():
+                keywords = recommended_kw[url]
+                if keywords != None:
+                    return None, None, gr.Radio(choices=keywords[:3], label="Recommended Keywords", visible=True,
+                                                interactive=True)
 
+            return None, None, gr.Radio(visible=False)
+
+
+        ctrlsum_file.upload(clear, inputs=None,
+                            outputs=[ctrlsum_str, ctrlsum_file_length, ctrlsum_file_keywords, ctrlsum_url, kw])
+        ctrlsum_url.input(update_url, inputs=ctrlsum_url, outputs=[ctrlsum_str, ctrlsum_file, kw])
 
-        ctrlsum_file.upload(clear, inputs=None,outputs=[ctrlsum_str,ctrlsum_file_length,ctrlsum_file_keywords, ctrlsum_url])
-        ctrlsum_url.input(clear, inputs=None, outputs=[ctrlsum_str, ctrlsum_file_length, ctrlsum_file_keywords, ctrlsum_file])
         ctrlsum_str.input(clear, inputs=None,
-                          outputs=[ctrlsum_url, ctrlsum_file_length, ctrlsum_file_keywords, ctrlsum_file])
+                          outputs=[ctrlsum_url, ctrlsum_file_length, ctrlsum_file_keywords, ctrlsum_file, kw])
+
+
+
+        def select_kw(env: gr.SelectData):
+            return env.value
+
+
+        kw.select(select_kw, None, ctrlsum_file_keywords)
+
         # Reference String Parsing
         with gr.TabItem("Reference String Parsing"):
             gr.Markdown(rsp_title_md)

controlled_summarization.py

@@ -3,9 +3,64 @@ import torch
 from SciAssist import Summarization
 import os
 import requests
+from datasets import load_dataset
+
+acl_data = load_dataset("dyxohjl666/CocoScisum_ACL", revision="refs/convert/parquet")
 device = "gpu" if torch.cuda.is_available() else "cpu"
 
-ctrlsum_pipeline = Summarization(os_name="nt",checkpoint="google/flan-t5-base",device=device)
+ctrlsum_pipeline = Summarization(os_name="nt",device=device)
+
+acl_dict = {}
+recommended_kw = {}
+
+
+def convert_to_dict(data):
+    """ Dict:
+        { url:
+            {length:
+                {keywords: summary};
+             raw_text:
+                 str;
+            }
+        }
+
+    """
+    url = data["url"]
+    text = data["text"]
+    keywords = data["keywords"]
+    length = data["length"]
+    summary = data["summary"]
+    for u, t, k, l, s in zip(url, text, keywords, length, summary):
+        if len(u) < 5:
+            continue
+        u = u + ".pdf"
+        if k == None:
+            k = ""
+        if l == None:
+            l = ""
+        k = str(k).strip()
+        l = str(l).strip()
+        if u in acl_dict.keys():
+            if k in acl_dict[u][l].keys():
+                continue
+            else:
+                acl_dict[u][l][k] = s
+        else:
+            acl_dict[u] = {"": {}, "50": {}, "100": {}, "200": {}, "raw_text": t}
+
+        # kws
+        if u in recommended_kw.keys():
+            if k == "" or k in recommended_kw[u]:
+                continue
+            else:
+                recommended_kw[u].append(k)
+        else:
+            recommended_kw[u] = []
+    return 1
+
+
+for i in acl_data.keys():
+    signal = convert_to_dict(acl_data[i])
 
 
 def download_pdf(url, dest_folder):
@@ -30,16 +85,15 @@ def download_pdf(url, dest_folder):
     return filename
 
 
-def ctrlsum_for_str(input,length=None, keywords=None) -> List[Tuple[str, str]]:
-
+def ctrlsum_for_str(input, length=None, keywords=None) -> List[Tuple[str, str]]:
     if keywords is not None:
         keywords = keywords.strip().split(",")
         if keywords[0] == "":
             keywords = None
-    if length==0 or length is None:
+    if length == 0 or length is None:
         length = None
     results = ctrlsum_pipeline.predict(input, type="str",
-                                    length=length, keywords=keywords)
+                                       length=length, keywords=keywords, num_beams=1)
 
     output = []
     for res in results["summary"]:
@@ -49,31 +103,49 @@ def ctrlsum_for_str(input,length=None, keywords=None) -> List[Tuple[str, str]]:
 
 def ctrlsum_for_file(input=None, length=None, keywords="", text="", url="") -> List[Tuple[str, str, str]]:
     if input == None and url == "":
-        if text=="":
-            return None,"Input cannot be left blank.",None
+        if text == "":
+            return None, "Input cannot be left blank.", None
         else:
-            return ctrlsum_for_str(text,length,keywords),text, None
+            return ctrlsum_for_str(text, length, keywords), text, None
     else:
-        filename=""
+        filename = ""
+        url = url.strip()
         if url != "":
-            if len(url) > 4:
+            if len(url) > 4 and url[-3:] == "pdf":
+                if url.strip() in acl_dict.keys():
+                    raw_text = acl_dict[url]["raw_text"]
+                    l = str(length)
+                    if length == 0:
+                        l = ""
+                    if l in acl_dict[url].keys():
+                        if keywords.strip() in acl_dict[url][l].keys():
+                            summary = acl_dict[url][l][keywords]
+                            return summary, raw_text, None
+                    if keywords.strip() == "":
+                        keywords = None
+                    if l == "":
+                        l = None
+                    return ctrlsum_for_str(raw_text, l, keywords), raw_text, None
+
                 filename = download_pdf(url, './cache/')
+            else:
+                "Invalid url(Not PDF)!", None, None
         else:
             filename = input.name
         if keywords != "":
             keywords = keywords.strip().split(",")
             if keywords[0] == "":
                 keywords = None
-        if length==0:
+        if length == 0:
             length = None
         # Identify the format of input and parse reference strings
         if filename[-4:] == ".txt":
             results = ctrlsum_pipeline.predict(filename, type="txt",
-                                            save_results=False,
-                                            length=length, keywords=keywords)
+                                               save_results=False,
+                                               length=length, keywords=keywords, num_beams=1)
         elif filename[-4:] == ".pdf":
             results = ctrlsum_pipeline.predict(filename,
-                                            save_results=False, length=length, keywords=keywords)
+                                               save_results=False, length=length, keywords=keywords, num_beams=1)
         else:
             return "File Format Error !", None, filename
 
@@ -83,5 +155,4 @@ def ctrlsum_for_file(input=None, length=None, keywords="", text="", url="") -> L
         return "".join(output), results["raw_text"], filename
 
 
-
-ctrlsum_str_example = "Language model pre-training has been shown to be effective for improving many natural language processing tasks ( Dai and Le , 2015 ; Peters et al. , 2018a ; Radford et al. , 2018 ; Howard and Ruder , 2018 ) . These include sentence-level tasks such as natural language inference ( Bowman et al. , 2015 ; Williams et al. , 2018 ) and paraphrasing ( Dolan and Brockett , 2005 ) , which aim to predict the relationships between sentences by analyzing them holistically , as well as token-level tasks such as named entity recognition and question answering , where models are required to produce fine-grained output at the token level ( Tjong Kim Sang and De Meulder , 2003 ; Rajpurkar et al. , 2016 ) . There are two existing strategies for applying pre-trained language representations to downstream tasks : feature-based and fine-tuning . The feature-based approach , such as ELMo ( Peters et al. , 2018a ) , uses task-specific architectures that include the pre-trained representations as additional features . The fine-tuning approach , such as the Generative Pre-trained Transformer ( OpenAI GPT ) ( Radford et al. , 2018 ) , introduces minimal task-specific parameters , and is trained on the downstream tasks by simply fine-tuning all pretrained parameters . The two approaches share the same objective function during pre-training , where they use unidirectional language models to learn general language representations . We argue that current techniques restrict the power of the pre-trained representations , especially for the fine-tuning approaches . The major limitation is that standard language models are unidirectional , and this limits the choice of architectures that can be used during pre-training . For example , in OpenAI GPT , the authors use a left-toright architecture , where every token can only attend to previous tokens in the self-attention layers of the Transformer ( Vaswani et al. , 2017 ) . Such restrictions are sub-optimal for sentence-level tasks , and could be very harmful when applying finetuning based approaches to token-level tasks such as question answering , where it is crucial to incorporate context from both directions . In this paper , we improve the fine-tuning based approaches by proposing BERT : Bidirectional Encoder Representations from Transformers . BERT alleviates the previously mentioned unidirectionality constraint by using a `` masked language model '' ( MLM ) pre-training objective , inspired by the Cloze task ( Taylor , 1953 ) . The masked language model randomly masks some of the tokens from the input , and the objective is to predict the original vocabulary id of the masked arXiv:1810.04805v2 [ cs.CL ] 24 May 2019 word based only on its context . Unlike left-toright language model pre-training , the MLM objective enables the representation to fuse the left and the right context , which allows us to pretrain a deep bidirectional Transformer . In addition to the masked language model , we also use a `` next sentence prediction '' task that jointly pretrains text-pair representations . The contributions of our paper are as follows : • We demonstrate the importance of bidirectional pre-training for language representations . Unlike Radford et al . ( 2018 ) , which uses unidirectional language models for pre-training , BERT uses masked language models to enable pretrained deep bidirectional representations . This is also in contrast to Peters et al . ( 2018a ) , which uses a shallow concatenation of independently trained left-to-right and right-to-left LMs . • We show that pre-trained representations reduce the need for many heavily-engineered taskspecific architectures . BERT is the first finetuning based representation model that achieves state-of-the-art performance on a large suite of sentence-level and token-level tasks , outperforming many task-specific architectures . • BERT advances the state of the art for eleven NLP tasks . The code and pre-trained models are available at https : //github.com/ google-research/bert . "
+ctrlsum_str_example = "Language model pre-training has been shown to be effective for improving many natural language processing tasks ( Dai and Le , 2015 ; Peters et al. , 2018a ; Radford et al. , 2018 ; Howard and Ruder , 2018 ) . These include sentence-level tasks such as natural language inference ( Bowman et al. , 2015 ; Williams et al. , 2018 ) and paraphrasing ( Dolan and Brockett , 2005 ) , which aim to predict the relationships between sentences by analyzing them holistically , as well as token-level tasks such as named entity recognition and question answering , where models are required to produce fine-grained output at the token level ( Tjong Kim Sang and De Meulder , 2003 ; Rajpurkar et al. , 2016 ) . There are two existing strategies for applying pre-trained language representations to downstream tasks : feature-based and fine-tuning . The feature-based approach , such as ELMo ( Peters et al. , 2018a ) , uses task-specific architectures that include the pre-trained representations as additional features . The fine-tuning approach , such as the Generative Pre-trained Transformer ( OpenAI GPT ) ( Radford et al. , 2018 ) , introduces minimal task-specific parameters , and is trained on the downstream tasks by simply fine-tuning all pretrained parameters . The two approaches share the same objective function during pre-training , where they use unidirectional language models to learn general language representations . We argue that current techniques restrict the power of the pre-trained representations , especially for the fine-tuning approaches . The major limitation is that standard language models are unidirectional , and this limits the choice of architectures that can be used during pre-training . For example , in OpenAI GPT , the authors use a left-toright architecture , where every token can only attend to previous tokens in the self-attention layers of the Transformer ( Vaswani et al. , 2017 ) . Such restrictions are sub-optimal for sentence-level tasks , and could be very harmful when applying finetuning based approaches to token-level tasks such as question answering , where it is crucial to incorporate context from both directions . In this paper , we improve the fine-tuning based approaches by proposing BERT : Bidirectional Encoder Representations from Transformers . BERT alleviates the previously mentioned unidirectionality constraint by using a `` masked language model '' ( MLM ) pre-training objective , inspired by the Cloze task ( Taylor , 1953 ) . The masked language model randomly masks some of the tokens from the input , and the objective is to predict the original vocabulary id of the masked arXiv:1810.04805v2 [ cs.CL ] 24 May 2019 word based only on its context . Unlike left-toright language model pre-training , the MLM objective enables the representation to fuse the left and the right context , which allows us to pretrain a deep bidirectional Transformer . In addition to the masked language model , we also use a `` next sentence prediction '' task that jointly pretrains text-pair representations . The contributions of our paper are as follows : • We demonstrate the importance of bidirectional pre-training for language representations . Unlike Radford et al . ( 2018 ) , which uses unidirectional language models for pre-training , BERT uses masked language models to enable pretrained deep bidirectional representations . This is also in contrast to Peters et al . ( 2018a ) , which uses a shallow concatenation of independently trained left-to-right and right-to-left LMs . • We show that pre-trained representations reduce the need for many heavily-engineered taskspecific architectures . BERT is the first finetuning based representation model that achieves state-of-the-art performance on a large suite of sentence-level and token-level tasks , outperforming many task-specific architectures . • BERT advances the state of the art for eleven NLP tasks . The code and pre-trained models are available at https : //github.com/ google-research/bert . "

requirements.txt

@@ -1,4 +1,5 @@
 pip==23.2.1
 torch==1.12.0
-SciAssist==0.0.41
+SciAssist==0.1.3
 nltk~=3.7
+pytest