Systems and methods for multi-modal conversation summarization on a conversation platform
Embodiments described herein provide a multi-modal search-and-summarize tool for message platforms. Specifically, the multi-modal search-and-summarize tool may monitor conversational content of different formats, e.g., text, image, video, etc., and use multi-modal summarization models to generate a summary of the conversation channel. The summarization may be conducted via a search-and-summarize process in response to a specific user query, e.g., a user may enter “what did John and Josh say about the presentation tomorrow?” The multi-modal summarization model would first search for relevant conversation messages between user John and user Josh, identify communication files of different format (e.g., text messages, emojis, multimedia attachments, etc.), and then input the communication files to respective text or image encoders to generate a summary of the communication content.
1. A method of multi-modal summarization of communication on a messaging platform, the method comprising:
receiving, via a user interface, a user request for summarizing communication messages relating to a topic between a first user and a second user on a messaging platform;
searching, via a search engine, for messages between the first user and the second user on the messaging platform;
filtering the messages based on the topic from the user request by predicting, via a topic classification model, whether the messages are related to the topic and excluding a subset of messages that are predicted to be unrelated to the topic;
generating, by a text encoder of a multi-modal summarization model, a text representation from an input sequence corresponding to textual content from the filtered messages;
generating, by an image encoder of the multi-modal summarization model, an image representation of visual features in a multimedia attachment file from the filtered messages;
generating, via a decoder of the multi-modal summarization model, a text summary summarizing both the filtered messages and the multimedia attachment file based on a combination of the text representation generated by the text encoder and the image representation generated by the image encoder,
wherein the text summary comprises a text that references a time and/or a sender of the multimedia attachment file generated from metadata of the multimedia attachment file; and
transmitting, via the user interface, the generated text summary relating to the topic in response to the user request.
2. The method of claim 1 , wherein the user request takes a form of a natural language question, and wherein the method further comprises:
identifying, by parsing the natural language question, the first user, the second user and the topic.
3. The method of claim 1 , wherein the user request further comprises one or more parameters entered via the user interface, and wherein the one or more parameters indicate any of:
a time range;
one or more conversation channels on the messaging platform; and
a file format.
4. The method of claim 1 , wherein the topic classification model is trained on a dataset of messages annotated with respective topic labels.
5. The method of claim 1 , wherein the multimedia attachment file includes any of a spreadsheet file, a presentation slide file, an audio file, a video file and an image file.
6. The method of claim 5 , further comprising:
converting the audio file into a text document; or
extracting one or more video frames from the video file.
7. The method of claim 1 , further comprising:
receiving, from the user interface, user feedback relating to the text summary;
generating an updated text summary according to the user feedback; and
incorporating the updated text summary into a training dataset for training the multi-modal summarization model.
8. The method of claim 1 , wherein the predicting, via the topic classification model, whether the messages are related to the topic comprises:
extracting one or more entities from the user request; and
predicting, one or more keywords related to the extracted one or more entities.
9. A system of multi-modal summarization of communication on a messaging platform, the system comprising:
a user interface that receives a user request for summarizing communication messages relating to a topic between a first user and a second user on a messaging platform;
a memory storing a plurality of processor-executable instructions; and
one or more processors executing the plurality of processor-executable instructions to perform operations comprising:
searching, via a search engine, for messages between the first user and the second user on the messaging platform;
filtering the messages based on the topic from the user request by predicting, via a topic classification model, whether the messages are related to the topic and excluding a subset of messages that are predicted to be unrelated to the topic;
generating, by a text encoder of a multi-modal summarization model, a text representation from an input sequence corresponding to textual content from the filtered messages;
generating, by an image encoder of the multi-modal summarization model, an image representation of visual features in a multimedia attachment file from the filtered messages; and
generating, via a decoder of the multi-modal summarization model, a text summary summarizing both the filtered messages and the multimedia attachment file based on a combination of the text representation generated by the text encoder and the image representation generated by the image encoder, wherein the text summary comprises a text that references a time and/or a sender of the multimedia attachment file generated from metadata of the multimedia attachment file;
wherein the user interface transmits the generated text summary relating to the topic in response to the user request.
10. The system of claim 9 , wherein the user request takes a form of a natural language question, and wherein the operations further comprise:
identifying, by parsing the natural language question, the first user, the second user and the topic.
11. The system of claim 9 , wherein the user request further comprises one or more parameters entered via the user interface, and wherein the one or more parameters indicate any of:
a time range;
one or more conversation channels on the messaging platform; and
a file format.
12. The system of claim 9 , wherein the topic classification model is trained on a dataset of messages annotated with respective topic labels.
13. The system of claim 9 , wherein the multimedia attachment file includes any of a spreadsheet file, a presentation slide file, an audio file, a video file and an image file.
14. The system of claim 13 , wherein the operations further comprise:
converting the audio file into a text document; or
extracting one or more video frames from the video file.
15. A non-transitory processor-readable storage medium storing a plurality of processor-executed instructions for multi-modal summarization of communication on a messaging platform, the instructions being executed by one or more processors to perform operations comprising:
receiving, via a user interface, a user request for summarizing communication messages relating to a topic between a first user and a second user on a messaging platform;
searching, via a search engine, for messages between the first user and the second user on the messaging platform;
filtering the messages based on the topic from the user request by predicting, via a topic classification model, whether the messages are related to the topic and excluding a subset of messages that are predicted to be unrelated to the topic;
generating, by a text encoder of a multi-modal summarization model, a text representation from an input sequence corresponding to textual content from the filtered messages;
generating, by an image encoder of the multi-modal summarization model, an image representation of visual features in a multimedia attachment file from the filtered messages;
generating, via a decoder of the multi-modal summarization model, a text summary summarizing both the filtered messages and the multimedia attachment file based on a combination of the text representation generated by the text encoder and the image representation generated by the image encoder, wherein the text summary comprises a text that references a time and/or a sender of the multimedia attachment file generated from metadata of the multimedia attachment file; and
transmitting, via the user interface, the generated text summary relating to the topic in response to the user request.