Attributes quotes in a text to their speakers. This annotator uses a two-stage linking strategy in which quotes are first linked to mentions, then mentions are linked to speakers. In the end, each quote will be annotated with both a mention in the text and a speaker entity.
||Annotator class name
||MentionAnnotation, MentionBeginAnnotation, MentionEndAnnotation, MentionTypeAnnotation, MentionSieveAnnotation, SpeakerAnnotation, SpeakerSieveAnnotation, ParagraphIndexAnnotation
Quote to mention linking can be done using the following sieves:
||Trigram detection such as QUOTE-CHARACTER-VERB
||Dependency parse information to find the nsubj relation of speech verbs like “say”
||If there is only a single mention in non-quote text in the same paragraph as the quote, use that mention
||Links a quote to vocatives in a quote preceding or following the target quote
||If a quote appears at the end of a paragraph, link it to the final mention in the previous sentence
||Conversational sieve links the target quote to the same mention as the quote two before it if they appear to be in a conversation
||Supervised classifier sieve uses the trained model to link quotes to mentions
||Loose conversational sieve is similar to the conversational sieve but relaxes the requirements for how to determine if two quotes are considered subsequent turns in a conversation
Mention to Speaker linking can be done using the following sieves:
||Deterministicly links mentions to speakers using extact name match and coreference information
||Uses loose conversational patterns to link mentions to the same speakers as corresponds to the conversation
||Links the mention to the top speaker that matches the mention in gender in a window around the target quote, containing logic to deal with conversations and vocative mentions
||Links the mention to the majority speaker
- quoteattribution.charactersPath (required): path to file containing the character names, aliases, and gender information.
- quoteattribution.booknlpCoref (required): path to tokens file generated from book-nlp containing coref information.
- quoteattribution.QMSieves: list of sieves to use in the quote to mention linking phase (default=tri,dep,onename,voc,paraend,conv,sup,loose).
- quoteattribution.MSSieves: list of sieves to use in the mention to speaker linking phase (default=det,top).
- quoteattribution.model: path to trained model file.
- quoteattribution.familyWordsFile: path to file with family words list.
- quoteattribution.animacyWordsFile: path to file with animacy words list.
- quoteattribution.genderNamesFile: path to file with names list with gender information.
Citing Stanford Quote Attribution
Grace Muzny, Michael Fang, Angel X. Chang and Dan Jurafsky. 2017. A Two-stage Sieve Approach for Quote Attribution. Proceedings of the European Chapter of the Association for Computational Linguistics (EACL). [pdf] [bib]