Wizard of Wikipedia, knowledge source page title. (#3845)

parlai/tasks/wizard_of_wikipedia/agents.py

@@ -24,6 +24,7 @@
 import copy
 from parlai.core.teachers import FixedDialogTeacher, MultiTaskTeacher
 from parlai.utils.io import PathManager
+from parlai.utils import logging
 from parlai.utils.misc import warn_once
 from .build import build
 
@@ -757,6 +758,113 @@ def get(self, episode_idx, entry_idx=0):
         return a
 
 
+class WikiPageTitleTeacher(WizardDialogKnowledgeTeacher):
+    """
+    Generates the title of Wikipedia page used as source of knowledge.
+
+    The context provided by this teacher (`text`) is the conversation history, with chosen topic removed.
+    The label is the title of the Wikipedia page of the passage that wizard selected for crafting
+    the next utterance; in other words, the source of knowledge for this utterance.
+    """
+
+    def __init__(self, opt, shared=None):
+        self.opt = copy.deepcopy(opt)
+        self.opt['label_type'] = 'response'
+        super().__init__(self.opt, shared=shared)
+        self.id = 'WikiPageTitleTeacher'
+        self._conv_history_len = self.opt['conversation_history_length']
+        if not (self._conv_history_len > 0 or self._conv_history_len == -1):
+            logging.warning(
+                f'"{self._conv_history_len}" is an invalid value for --conversation-history-length flag.'
+                ' Changing it to default of -1 (include the entire message history).'
+            )
+            self._conv_history_len = -1
+        self._skip_no_title = self.opt['skip_no_title']
+        if not shared:
+            self._preprocess_data()
+        else:
+            self.titles_data = shared['titles_data']
+
+    @classmethod
+    def add_cmdline_args(cls, parser, partial_opt=None):
+        super().add_cmdline_args(parser, partial_opt=partial_opt)
+        agent = parser.add_argument_group('Wikipedia Page Title Arguments')
+        agent.add_argument(
+            '--conversation-history-length',
+            type=int,
+            default=-1,
+            help='Number of previous utterances to keep in context, 0 (default) includes all',
+        )
+        agent.add_argument(
+            '--skip-no-title',
+            type='bool',
+            default=True,
+            help=(
+                'Whether to skip the example if no passage was selected. If `false` '
+                f'uses `{TOKEN_NOCHOSEN}` instead of title if no knowledge source was selected.'
+            ),
+        )
+        return parser
+
+    def share(self):
+        shared = super().share()
+        shared['titles_data'] = self.titles_data
+        return shared
+
+    def _generate_messages(self, hist, action):
+        include_hist = (
+            hist[-self._conv_history_len :] if self._conv_history_len > 0 else hist
+        )
+        context = '\n'.join(include_hist)
+        return Message(
+            {
+                'id': "Wikipedia Title Teacher",
+                'text': context,
+                'labels': [action["title"]],
+                'episode_done': True,
+            }
+        )
+
+    def _should_include(self, act):
+        return not (self._skip_no_title and act['labels'][0] == TOKEN_NOCHOSEN)
+
+    def _preprocess_data(self):
+        data = []
+        for episode_idx in range(super().num_episodes()):
+            dialog_history = []
+            ex_idx = 0
+            while True:
+                a = super().get(episode_idx, ex_idx)
+                text_parts = a['text'].split('\n')
+                if ex_idx == 0:
+                    # throwing away chosen_topic
+                    text_parts = text_parts[1:]
+                if text_parts:
+                    dialog_history.append(text_parts[0])
+                    title_act = self._generate_messages(dialog_history, a)
+                    if self._should_include(title_act):
+                        data.append(title_act)
+                if a['episode_done']:
+                    break
+                ex_idx += 1
+                dialog_history.append(a['labels'][0])
+
+        logging.info(
+            f'{len(data)} title generation examples generated '
+            f'from {super().num_examples()} original examples'
+        )
+        self.titles_data = data
+
+    def num_episodes(self):
+        return len(self.titles_data)
+
+    def num_examples(self):
+        return self.num_episodes()
+
+    def get(self, episode_idx, entry_idx=0):
+        return self.titles_data[episode_idx]
+
+
 ####################################################
 #                                                  #
 # Doc Reader Teachers                              #

parlai/tasks/wizard_of_wikipedia/test.py

@@ -39,5 +39,9 @@ class TestGeneratorTeacher(AutoTeacherTest):
     task = "wizard_of_wikipedia:generator"
 
 
+class TestWikiPageTitleTeacher(AutoTeacherTest):
+    task = "wizard_of_wikipedia:wiki_page_title"
+
+
 class TestDocreaderTeacher(AutoTeacherTest):
     task = "wizard_of_wikipedia:docreader"

parlai/tasks/wizard_of_wikipedia/test/wizard_of_wikipedia_wiki_page_title_test.yml

@@ -0,0 +1,46 @@
+acts:
+- - episode_done: true
+    eval_labels:
+    - Royal Blue (train)
+    id: WikiPageTitleTeacher
+    text: 'Blue is my favorite primary color.
+
+      Blue is always nice. I like royal blue.'
+- - episode_done: true
+    eval_labels:
+    - Blue Skies (1946 film)
+    id: WikiPageTitleTeacher
+    text: 'Blue is my favorite primary color.
+
+      Blue is always nice. I like royal blue.
+
+      I once road on The Royal Blue train from New York to D.C
+
+      Oh that sounds really nice. I bet there was a lot of scenery and blue skies.'
+- - episode_done: true
+    eval_labels:
+    - Cinematography
+    id: WikiPageTitleTeacher
+    text: Hi buddy, What you think about cinematography
+- - episode_done: true
+    eval_labels:
+    - Cinematography
+    id: WikiPageTitleTeacher
+    text: "Hi buddy, What you think about cinematography\nCinematography,is a type\
+      \ of motion picture , captured electronically by means of an image \nYes buddy,\
+      \  Images captured with an electronic image-sensor, produces an electrical charge.The\
+      \ word \"cinematography\" is based on the Greek words  meaning movement, motion."
+- - episode_done: true
+    eval_labels:
+    - Photography
+    id: WikiPageTitleTeacher
+    text: "Hi buddy, What you think about cinematography\nCinematography,is a type\
+      \ of motion picture , captured electronically by means of an image \nYes buddy,\
+      \  Images captured with an electronic image-sensor, produces an electrical charge.The\
+      \ word \"cinematography\" is based on the Greek words  meaning movement, motion.\n\
+      It works by lens used to repeatedly focus the light reflected from objects into\
+      \ real images on the light-sensitive surface .\n Muybridge sequence of a horse\
+      \ galloping In the 1830s, moving images were produced on revolving drums and\
+      \ disks"
+num_episodes: 3181
+num_examples: 3181

parlai/tasks/wizard_of_wikipedia/test/wizard_of_wikipedia_wiki_page_title_train.yml

@@ -0,0 +1,48 @@
+acts:
+- - episode_done: true
+    id: WikiPageTitleTeacher
+    labels:
+    - Science fiction film
+    text: 'I think science fiction is an amazing genre for anything. Future science,
+      technology, time travel, FTL travel, they''re all such interesting concepts.
+
+      I''m a huge fan of science fiction myself! '
+- - episode_done: true
+    id: WikiPageTitleTeacher
+    labels:
+    - Time travel in fiction
+    text: "I think science fiction is an amazing genre for anything. Future science,\
+      \ technology, time travel, FTL travel, they're all such interesting concepts.\n\
+      I'm a huge fan of science fiction myself! \nAwesome! I really love how sci-fi\
+      \ storytellers focus on political/social/philosophical issues that would still\
+      \ be around even in the future. Makes them relatable.\nI agree. One of my favorite\
+      \ forms of science fiction is anything related to time travel! I find it fascinating."
+- - episode_done: true
+    id: WikiPageTitleTeacher
+    labels:
+    - Science fiction
+    text: "I think science fiction is an amazing genre for anything. Future science,\
+      \ technology, time travel, FTL travel, they're all such interesting concepts.\n\
+      I'm a huge fan of science fiction myself! \nAwesome! I really love how sci-fi\
+      \ storytellers focus on political/social/philosophical issues that would still\
+      \ be around even in the future. Makes them relatable.\nI agree. One of my favorite\
+      \ forms of science fiction is anything related to time travel! I find it fascinating.\n\
+      It's not quite sci-fi, but my favorite version of time travel is in Harry Potter\
+      \ and the Prisoner of Azkaban. Breaks zero logical rules.\nAnd that's difficult\
+      \ to do when dealing with time travel. I actually haven't seen the latest Harry\
+      \ Potter movies. Guess it's time to check them out!"
+- - episode_done: true
+    id: WikiPageTitleTeacher
+    labels:
+    - Internet access
+    text: 'Can you imagine the world without internet access? '
+- - episode_done: true
+    id: WikiPageTitleTeacher
+    labels:
+    - Internet access
+    text: "Can you imagine the world without internet access? \nNo I could not! I\
+      \ couldn't imagine living when internet access was rare and very few people\
+      \ had it!\nOh me either! It seems like such a long time ago. I wonder when Internet\
+      \ was first created?"
+num_episodes: 60797
+num_examples: 60797

parlai/tasks/wizard_of_wikipedia/test/wizard_of_wikipedia_wiki_page_title_valid.yml

@@ -0,0 +1,61 @@
+acts:
+- - episode_done: true
+    eval_labels:
+    - Gardening
+    id: WikiPageTitleTeacher
+    text: I like Gardening, even when I've only been doing it for a short time.
+- - episode_done: true
+    eval_labels:
+    - Gardening
+    id: WikiPageTitleTeacher
+    text: 'I like Gardening, even when I''ve only been doing it for a short time.
+
+      I live on a farm, we garden all year long, it is very relaxing.
+
+      That sounds great.  I''ve always thought that I would love living in a farm,
+      but I;ve always lived in the city.  What do you mostly plant?'
+- - episode_done: true
+    eval_labels:
+    - Gardening
+    id: WikiPageTitleTeacher
+    text: 'I like Gardening, even when I''ve only been doing it for a short time.
+
+      I live on a farm, we garden all year long, it is very relaxing.
+
+      That sounds great.  I''ve always thought that I would love living in a farm,
+      but I;ve always lived in the city.  What do you mostly plant?
+
+      I have planted several fruits trees, tomatoes, jalepenos, bell peppers, onions,
+      Garlic, and potatoes mostly.
+
+      Great, I love the idea of growing my own vegetables and fruits! Do you have
+      animals in the farm?'
+- - episode_done: true
+    eval_labels:
+    - Gardening
+    id: WikiPageTitleTeacher
+    text: 'I like Gardening, even when I''ve only been doing it for a short time.
+
+      I live on a farm, we garden all year long, it is very relaxing.
+
+      That sounds great.  I''ve always thought that I would love living in a farm,
+      but I;ve always lived in the city.  What do you mostly plant?
+
+      I have planted several fruits trees, tomatoes, jalepenos, bell peppers, onions,
+      Garlic, and potatoes mostly.
+
+      Great, I love the idea of growing my own vegetables and fruits! Do you have
+      animals in the farm?
+
+      yes i do. Cows, chickens, Micro pigs, Guinneas, We also do forest growing also.
+      we plants large pine trees.
+
+      Wow, it sounds amazing, the Micro-pigs are so cute! are they trainable to be
+      well behaved?'
+- - episode_done: true
+    eval_labels:
+    - Bob Ross
+    id: WikiPageTitleTeacher
+    text: I would like to know more about bob ross
+num_episodes: 3236
+num_examples: 3236