Extraction of tasks from documents using weakly supervision
This disclosure relates to extraction of tasks from documents based on a weakly supervised classification technique, wherein extraction of tasks is identification of mentions of tasks in a document. There are several prior arts addressing the problem of extraction of events, however due to crucial distinctions between events-tasks, task extraction stands as a separate problem. The disclosure explicitly defines specific characteristics of tasks, creates labelled data at a word-level based on a plurality of linguistic rules to train a word-level weakly supervised model for task extraction. The labelled data is created based on the plurality of linguistic rules for a non-negation aspect, a volitionality aspect, an expertise aspect and a plurality of generic aspects. Further the disclosure also includes a phrase expansion technique to capture the complete meaning expressed by the task instead of merely mentioning the task that may not capture the entire meaning of the sentence.
1. A processor-implemented method for training a word-level data model for extraction of tasks from documents using weakly supervision comprising:
receiving a plurality of documents from a plurality of sources, via one or more hardware processors;
pre-processing the plurality of documents using a plurality of pre-processing techniques, via the one or more hardware processors, to obtain a plurality of pre-processed documents comprising a plurality of sentences, a plurality of words within the plurality of sentences, a plurality of Part-of-speech (POS) tags, a plurality of dependency trees and a plurality of WordNet® based features;
labelling the plurality of words from the plurality of pre-processed documents as one of a task headword and a no-task headword, via the one or more hardware processors, wherein the plurality of words is labelled based on the plurality of sentences using a plurality of linguistic rules, wherein the plurality of words is labelled based on the plurality of linguistic rules for a volitionality aspect referring to tasks whose actions are carried out by an actor volitionally,
wherein the volitionality aspect includes an action verbs or nouns aspect, an animate organization agent aspect, an inanimate agent aspect and a volition marker,
wherein the animate organization agent aspect captures the volitionality in an implicit way as animate agents indicate that the action corresponding to verb w is likely to be carried out volitionally,
wherein for the animate organization agent aspect, when the agent of the verb w is animate or corresponds to an organization, then the word is labelled as “task headword”, wherein for the inanimate agent aspect, when the agent of the verb w is inanimate, then the word is labelled as “no task headword”; and
training a word-level weakly supervised classification model for extraction of tasks, via the one or more hardware processors, using the task headword and the no-task headword labelled using the plurality of linguistic rules, wherein the word-level weakly supervised classification model is the word-level data model for extraction of tasks from documents, wherein the word-level weakly supervised classification model is a Bidirectional Encoder Representations from Transformers (BERT-based classification model), wherein each instance is annotated with soft label and each instance is a combination of a word w, the word's POS tag p, and the sentence S.
2. The method of claim 1 , wherein the word-level weakly supervised model is utilized for the extraction of tasks from a plurality of user documents based on a word-level weakly supervision task extraction technique, wherein the weakly supervision task extraction technique comprises:
receiving the plurality of user documents for the extraction of tasks;
pre-processing the plurality of user documents using the plurality of pre-processing techniques, to obtain a plurality of pre-processed user documents comprising a plurality of user sentences and a plurality of user words within the plurality of user sentences;
labelling the plurality of user words from the plurality of pre-processed user documents as one of a task headword and a no-task headword using the word-level weakly supervised classification model; and
expanding the labelled user task headword to obtain a task phrase based a phrase expansion technique and the plurality of dependency trees, wherein the task phrase represents the extracted task.
3. The method of claim 1 , wherein the task extraction comprises identification of mentions of a task in a document, wherein a task is a well-defined knowledge-based action carried out volitionally with expertise for a specific goal within a pre-defined time by a single person, a group of persons, a device, or a system.
4. The method of claim 1 , wherein the plurality of pre-processing techniques comprises (a) a sentence-splitting technique for identification of the plurality of sentences, (b) a tokenization technique for identification of the plurality of words within the plurality of sentences, (c) a Part of speech (POS) tagging technique for identification of the plurality of Part-of-speech (POS) tags for each word in the plurality of words, (d) a dependency parsing technique for identification of the plurality of dependency trees for the plurality of sentences, and (e) a WordNet® based features identification technique for identification of the plurality of WordNet® based features for the plurality of sentences.
5. The method of claim 1 , wherein the plurality of words is labelled based on the plurality of linguistic rules for a non-negation aspect, an expertise aspect and a plurality of generic aspects.
6. The method of claim 5 , wherein the plurality of linguistic rules for labelling the non-negation aspect labels the plurality of words based on a modified negation using the plurality of dependency trees, the plurality of linguistic rules for labelling the volitionality aspect labels the plurality of words based on identification of actions carried out volitionally, and the plurality of linguistic rules for labelling the expertise aspect labels the plurality of words based on a domain expertise or knowledge required for execution of the task.
7. A system comprising:
a memory storing instructions;
one or more communication interfaces; and
one or more hardware processors coupled to the memory via the one or more communication interfaces, wherein the one or more hardware processors are configured by the instructions to:
receive a plurality of documents from a plurality of sources, via one or more hardware processors;
pre-process the plurality of documents using a plurality of pre-processing techniques, via the one or more hardware processors, to obtain a plurality of pre-processed documents comprising a plurality of sentences, a plurality of words within the plurality of sentences, a plurality of Part-of-speech (POS) tags, a plurality of dependency trees and a plurality of WordNet® based features;
label the plurality of words from the plurality of pre-processed documents as one of a task headword and a no-task headword, via the one or more hardware processors, wherein the plurality of words is labelled based on the plurality of sentences using a plurality of linguistic rules, wherein the plurality of words is labelled based on the plurality of linguistic rules for a volitionality aspect referring to tasks whose actions are carried out by an actor volitionally,
wherein the volitionality aspect includes an action verbs or nouns aspect, an animate organization agent aspect, an inanimate agent aspect and a volition marker,
wherein the animate organization agent aspect captures the volitionality in an implicit way as animate agents indicate that the action corresponding to verb w is likely to be carried out volitionally,
wherein for the animate organization agent aspect, when the agent of the verb w is animate or corresponds to an organization, then the word is labelled as “task headword”, wherein for the inanimate agent aspect, when the agent of the verb w is inanimate, then the word is labelled as “no task headword”; and
train a word-level weakly supervised classification model for extraction of tasks, via the one or more hardware processors, using the task headword and the no-task headword labelled using the plurality of linguistic rules, wherein the word-level weakly supervised classification model is the word-level data model for extraction of tasks from documents, wherein the word-level weakly supervised classification model is a Bidirectional Encoder Representations from Transformers (BERT-based classification model), and wherein each instance is annotated with soft labels, wherein each instance is a combination of a word w, the word's POS tag p, and the sentence S.
8. The system of claim 7 , wherein the one or more hardware processors are configured by the instructions to perform the extraction of tasks from a plurality of user documents using the word-level weakly supervised model based on a word-level weakly supervision task extraction technique, wherein the weakly supervision task extraction technique comprises:
receiving the plurality of user documents for the extraction of tasks;
pre-processing the plurality of user documents using the plurality of pre-processing techniques, to obtain a plurality of pre-processed user documents comprising a plurality of user sentences and a plurality of user words within the plurality of user sentences;
labelling the plurality of user words from the plurality of pre-processed user documents as one of a task headword and a no-task headword using the word-level weakly supervised classification model; and
expanding the labelled user task headword to obtain a task phrase based a phrase expansion technique and the plurality of dependency trees, wherein the task phrase represents the extracted task.
9. The system of claim 7 , wherein the one or more hardware processors are configured by the instructions to perform the task extraction comprising identification of mentions of a task in a document, wherein a task is a well-defined knowledge-based action carried out volitionally with expertise for a specific goal within a pre-defined time by a single person, a group of persons, a device, or a system.
10. The system of claim 7 , wherein the one or more hardware processors are configured by the instructions to perform the plurality of pre-processing technique comprising (a) a sentence-splitting technique for identification of the plurality of sentences, (b) a tokenization technique for identification of the plurality of words within the plurality of sentences, (c) a Part of speech (POS) tagging technique for identification of the plurality of Part-of-speech (POS) tags for each word in the plurality of words, (d) a dependency parsing technique for identification of the plurality of dependency tree for the plurality of sentences, and a WordNet® based features identification technique for identification of the plurality of WordNet® based features for the plurality of sentences.
11. The system of claim 7 , wherein the one or more hardware processors are configured by the instructions to perform labeling of the plurality of words based on the plurality of linguistic rules for a non-negation aspect, an expertise aspect and a plurality of generic aspects.
12. The system of claim 11 , wherein the one or more hardware processors are configured by the instructions to perform the labelling of the plurality of words based on the plurality of linguistic rules incudes (a) the non-negation aspect labeling the plurality of words based on a modified by a negation using the dependency tree, (b) the volitionality aspect labeling the plurality of words based on identification of actions carried out volitionally, and the (c) the expertise aspect labeling the plurality of words based on a domain expertise or knowledge required for execution of the task.
13. A computer program product comprising a non-transitory computer readable medium having a computer readable program embodied therein, wherein the computer readable program, when executed on a computing device, causes the computing device to:
receive a plurality of documents from a plurality of sources, via one or more hardware processors;
pre-process the plurality of documents using a plurality of pre-processing techniques, via the one or more hardware processors, to obtain a plurality of pre-processed documents comprising a plurality of sentences, a plurality of words within the plurality of sentences, a plurality of Part-of-speech (POS) tags, a plurality of dependency trees and a plurality of WordNet® based features;
label the plurality of words from the plurality of pre-processed documents as one of a task headword and a no-task headword, via the one or more hardware processors, wherein the plurality of words is labelled based on the plurality of sentences using a plurality of linguistic rules, wherein the plurality of words is labelled based on the plurality of linguistic rules for a volitionality aspect referring to tasks whose actions are carried out by an actor volitionally, wherein the volitionality aspect includes an action verbs or nouns aspect, an animate organization agent aspect, an inanimate agent aspect and a volition marker, wherein the animate organization agent aspect captures the volitionality in an implicit way as animate agents indicate that the action corresponding to verb w is likely to be carried out volitionally, wherein for the animate organization agent aspect, when the agent of the verb w is animate or corresponds to an organization, then the word is labelled as “task headword”, wherein for the inanimate agent aspect, when the agent of the verb w is inanimate, then the word is labelled as “no task headword”; and
train a word-level weakly supervised classification model for extraction of tasks, via the one or more hardware processors, using the task headword and the no-task headword labelled using the plurality of linguistic rules, wherein the word-level weakly supervised classification model is the word-level data model for extraction of tasks from documents, wherein the word-level weakly supervised classification model is a Bidirectional Encoder Representations from Transformers (BERT-based classification model), and wherein each instance is annotated with soft labels, wherein each instance is a combination of a word w, the word's POS tag p, and the sentence S.