Merge pull request #30 from nbcstevenchen/main

Create Semantic Ranking for Title, Summary, Keywords, Title and Conference Name Searching

cncf-youtube-channel-summarizer/embedding_generator.py

@@ -0,0 +1,18 @@
+from sentence_transformers import SentenceTransformer
+import pandas as pd
+import pickle
+def embedding_generator(model, data):
+    dataset = pd.read_csv('data/cncf_video_summary_combine.csv')
+    bi_encoder = SentenceTransformer(model)
+    bi_encoder.max_seq_length = 256     #Truncate long passages to 256 tokens
+    ##### Semantic Search #####
+    # Encode the query using the bi-encoder and find potentially relevant passages
+    embeddings = bi_encoder.encode(data, convert_to_tensor=True, show_progress_bar=True)
+    return embeddings
+
+
+if __name__ == "__main__":
+    dataset = pd.read_csv('data/cncf_video_summary_combine.csv')
+    embeddings = embedding_generator('multi-qa-MiniLM-L6-cos-v1', dataset['merge'])
+    with open('data/embedding.pkl', 'wb') as f:
+        pickle.dump(embeddings.numpy(), f)

cncf-youtube-channel-summarizer/merge_csv.py

@@ -0,0 +1,21 @@
+import pandas as pd
+import os
+
+def merge_csv():
+    csv_list = []
+    filename_list = os.listdir('data/')
+    for i in range(0, len(filename_list)):
+        if '.csv' in filename_list[i]:
+            filename_list[i] = 'data/' + filename_list[i]
+            csv_list.append(filename_list[i])
+
+    dataframes = [pd.read_csv(each_file) for each_file in csv_list]
+    merged_df = pd.concat(dataframes, ignore_index=True)
+    merged_df['merge'] = merged_df['video_title'] + ' ' + merged_df['conference_name'] + ' ' + merged_df[
+        'summary'] + ' ' + merged_df['keywords']
+    merged_df.to_csv('data/cncf_video_summary_combine.csv', index=False)
+
+if __name__ == "__main__":
+    merge_csv()
+
+

cncf-youtube-channel-summarizer/requirements.txt

@@ -6,3 +6,6 @@ tenacity==8.2.3
 tqdm==4.64.0
 youtube_transcript_api==0.6.2
 transformers==4.41.1
+nltk==3.8.1
+rank_bm25==0.2.2
+sentence_transformers==3.0.1

cncf-youtube-channel-summarizer/semantic_searching.py

@@ -0,0 +1,82 @@
+import nltk
+from nltk.tokenize import word_tokenize
+from rank_bm25 import BM25Okapi
+import pandas as pd
+from sentence_transformers import SentenceTransformer, util
+import pickle
+# Download required NLTK data
+nltk.download('punkt')
+
+class BM25():
+    def __init__(self, dataset, top_k=5):
+        self.dataset = dataset
+        self.top_k = top_k
+        self.tokenized_corpus = [self.preprocess_text(doc) for doc in dataset['merge']]
+    # Function to preprocess and tokenize text
+    def preprocess_text(self, text):
+        return word_tokenize(text.lower())
+
+
+    # Function to perform a search query
+    def search(self, query, bm25):
+        tokenized_query = self.preprocess_text(query)
+        scores = bm25.get_scores(tokenized_query)
+        results = []
+        top_n_indices = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:self.top_k]
+
+        video_ids = []
+        for i in top_n_indices:
+            results.append((self.dataset['merge'][i], scores[i]))
+            video_ids.append(self.dataset.loc[i]['video_id'])
+        print(results)
+        return video_ids
+
+
+    def run(self, query):
+        # Initialize BM25
+        bm25 = BM25Okapi(self.tokenized_corpus)
+        # Example query
+       # query = "CNCF Webinars"
+        video_ids = self.search(query, bm25, )
+        return video_ids
+
+class BIENCODER():
+    def __init__(self, dataset, embeddings, top_k=5):
+        self.dataset = dataset
+        self.embeddings = embeddings
+        self.top_k = top_k
+        self.bi_encoder = SentenceTransformer('multi-qa-MiniLM-L6-cos-v1')
+        self.bi_encoder.max_seq_length = 256
+
+    def search(self, query):
+        print("Input question:", query)
+        question_embedding = self.bi_encoder.encode(query, convert_to_tensor=True)
+       # question_embedding = question_embedding.cuda()
+        hits = util.semantic_search(question_embedding, self.embeddings, top_k=self.top_k)
+        hits = hits[0]  # Get the hits for the first query
+        # print(hits)
+
+        # Output of top-5 hits from bi-encoder
+        print("\n-------------------------\n")
+        print("Top-3 Bi-Encoder Retrieval hits")
+        hits = sorted(hits, key=lambda x: x['score'], reverse=True)
+        video_ids = []
+        for hit in hits:
+            print("\t{:.3f}\t{}".format(hit['score'], self.dataset['merge'][hit['corpus_id']]))
+            video_ids.append(self.dataset.loc[hit['corpus_id']]['video_id'])
+        return video_ids
+
+if __name__ == "__main__":
+    query = 'CNCF Webinars'  ## input query
+    dataset = pd.read_csv('data/cncf_video_summary_combine.csv')
+    print('Method 1: BM25 alg for semantic search:')
+    bm25_search = BM25(dataset, top_k=5)
+    video_ids = bm25_search.run(query)
+    print('here')
+    print(video_ids)
+
+    print('Method 2: Deep learning for semantic search:')
+    with open('data/embedding.pkl', 'rb') as f:
+        embeddings = pickle.load(f)
+    video_ids = BIENCODER(dataset, embeddings).search(query)
+    print(video_ids)