How to Build a Text Summarizer using Huggingface Transformers | Turingtalks — Transcript

Full transcript

1. 0:00hello guys welcome to this channel in
2. 0:02today's video we're going to see how to
3. 0:04build a tech summarizer using hugging
4. 0:06phas
5. 0:07Transformers so in this video I'll walk
6. 0:09you through what a summarizer is its use
7. 0:11cases what huging face Transformers are
8. 0:14and how you can build your own text
9. 0:15summarizer using huging face
10. 0:16Transformers Library so let's get
11. 0:19started a summarizer does exactly what
12. 0:21the name suggests it takes a large block
13. 0:24of text and condenses it into a shorter
14. 0:26version so this shorter version keeps
15. 0:28only the key points think think of it as
16. 0:30the difference between reading a whole
17. 0:31novel and glancing at its back cover so
18. 0:34the aim is to save time while still
19. 0:36getting the essence of the content now
20. 0:38where do we use the summarizer for
21. 0:40example journalists use them to quickly
22. 0:42sift through reports and studies
23. 0:44students can use them to summarize
24. 0:46lengthy readings businesses can use them
25. 0:48to condense market analysis or lengthy
26. 0:50reports in essence anyone who needs to
27. 0:53process large amounts of text can
28. 0:55quickly benefit from a
29. 0:57summarizer now what are hugin phace
30. 0:58Transformers hugging face is a company
31. 1:01that has created a state-of-the-art
32. 1:03platform for natural language processing
33. 1:05they have a library called the
34. 1:07Transformers Library it is like a
35. 1:09treasure toe for NLP tasks it includes
36. 1:11pre-train models that can do everything
37. 1:13from translation sentiment analysis and
38. 1:15yes summarization so these models have
39. 1:18learned from vast amounts of text and
40. 1:20can understand and generate language in
41. 1:21a surprisingly human-like way so we
42. 1:23don't have to do a lot of the training
43. 1:24in these models these are pre-trained
44. 1:26models that we can take them and even
45. 1:28fine-tune them to use it for our needs
46. 1:31now let's see how we can build a
47. 1:33summarizer using the hugging face
48. 1:34Transformers model for this example we
49. 1:37will be using Facebook's Bart model so
50. 1:39Bart model is a pre-trained model on the
51. 1:42English language so it is a sequence to
52. 1:44sequence model and it is great for uh
53. 1:45text generation for example
54. 1:47summarization translation but it also
55. 1:49works well for comprehension tasks like
56. 1:51text classification question answering
57. 1:53and so many others so a sequence to
58. 1:54sequence model is where you input a
59. 1:56sequence and it outputs a sequence for
60. 1:58example Chach PT Hing Fai also has a
61. 2:01concept called pipelines so pipelines
62. 2:03offer a simpler approach to implementing
63. 2:05various tasks so instead of preparing a
64. 2:07data set training it with the model and
65. 2:08then using it pipeline simplifies the
66. 2:10code because it HIDs away a lot of the
67. 2:13manual work uh like tokenization and
68. 2:15model customization and you can just get
69. 2:17started working with the model so we
70. 2:18will be using a Google collab notebook
71. 2:20for this project you can also find a
72. 2:22link for the finished project in the
73. 2:23description below so let's open the
74. 2:25notebook and let's first install
75. 2:27Transformers so installing Transformers
76. 2:29will be a sh command whereas colloud
77. 2:31notebooks read python so we will add an
78. 2:33exclamation mark at the beginning so pip
79. 2:35install Transformers let's wait for the
80. 2:38Transformers library to be installed
81. 2:40great Transformers library is now
82. 2:42installed so now let's initialize a text
83. 2:45summarization by bline using the huging
84. 2:47face Transformers Library here I'll be
85. 2:49typing summarizer equal to by blind
86. 2:51summarizer model Facebook but L CNN so
87. 2:54let's see what each part does the
88. 2:57pipeline is a function provided by the
89. 2:58huging face Transformers SL to make it
90. 3:00easy to apply different types of natural
91. 3:02language processing tasks so the
92. 3:04function returns a ready to use pipeline
93. 3:06object for the specified tasks
94. 3:08summarization is the first argument we
95. 3:09are telling the pipeline that task that
96. 3:11we are going to perform is summarization
97. 3:13and the model here is going to be
98. 3:15Facebook's bot so we'll be using
99. 3:17facebook/ bot CNN so this argument
100. 3:20specifies the pre-train model that will
101. 3:23be used for this project for this
102. 3:24summarization task is Facebook's B Cloud
103. 3:28CNN so this mod is provided by Facebook
104. 3:30and it is placed on the bot model bot
105. 3:32architecture which is effective for uh
106. 3:34General NLP tasks so let's execute the
107. 3:37code how creates a summarizer object so
108. 3:39this object can be used to perform text
109. 3:41summarization by passing Text data to it
110. 3:44the model will generate a shorter
111. 3:45version of the input text and it will
112. 3:47capture the most important or relevant
113. 3:49information according to its training on
114. 3:50the summarization tasks so yeah we are
115. 3:52now ready to use the model uh that's all
116. 3:55it takes to build a summarization model
117. 3:57using hugging the face pipeline so you
118. 3:58can see how easy it is to build build
119. 4:00NLP models using hugging face
120. 4:02Transformers so we will fetch an anal
121. 4:04report from FDA here it
122. 4:06is so let's copy the text and create a
123. 4:10variable called text and put in our text
124. 4:13now let's use this text as input and
125. 4:14call our summarizer here I'll be writing
126. 4:16summarizer equal to summarizer text Max
127. 4:18length Min length and do
128. 4:22sample this line of code is using the
129. 4:24summarizer object created from a hugging
130. 4:26face pipeline to generate a summary of
131. 4:28the input text so summarizer is the
132. 4:30object initialized previously with the
133. 4:32bip planine function the text we have
134. 4:34already declared we setting the Max and
135. 4:36Min length so the max length specifies
136. 4:39the maximum length uh obviously and uh
137. 4:42it'll be in terms of the number of
138. 4:43tokens it's not the number of characters
139. 4:45it's the number of tokens and the Min
140. 4:46length is also the minimum length of the
141. 4:48summary so we don't want the summary to
142. 4:50be too short so now we can go ahead and
143. 4:53print the summary and here's the
144. 4:55response so in short this code loads a
145. 4:57summarization pipeline that is
146. 4:58preconfigured to use the Facebook B
147. 5:00model feeds the text to A summarizer
148. 5:03outputs a summary with a specified
149. 5:04minimum and maximum length so you can
150. 5:06see how easy it is to work with the
151. 5:08hugging face Transformers Library there
152. 5:10are many other features or NLP functions
153. 5:13that are offered by hugging face
154. 5:14Transformers Library so I'll be posting
155. 5:16more videos on those topics if you want
156. 5:18to get daily AI tutorials we have a
157. 5:21newsletter called touring talks you can
158. 5:23find the newsletter at during talks. you
159. 5:26can also subscribe to this channel to
160. 5:28get notified when we post new videos
161. 5:30thanks for watching if you have any
162. 5:32questions please let us know in the
163. 5:33comments see you soon with a new topic
164. 5:35cheers