Description:Field of the Invention
The goal of this invention is to create algorithms and systems that automatically summarize business talks using NLP methods and AI technology. To make it simpler for participants to review and recall the crucial details without having to thoroughly comb through the meeting transcript or recording, these summaries strive to capture the essential ideas, conclusions, action items, and other pertinent information mentioned during the meeting. Typical NLP tasks include text summarizing, entity recognition, sentiment analysis, speech-to-text conversion, and possibly even some degree of machine learning to enhance the precision and quality of the generated summaries.
OBJECTIVE OF THIS INVENTION
The main goal is to develop an automated system that can produce succinct and educational summaries of business meetings. The invention aims to save participants' time by automatically reducing the information from protracted business meetings into summaries that are focused and concise. The meeting's major discussions, decisions, and action items should all be well summarized in the summaries that are produced. By giving a succinct and clear summary of the meeting talks, the invention aids in the making of informed decisions. As a result, participants may concentrate on the key points and base their decisions on the distilled knowledge in an educated manner.
Background of the Invention
The background of the invention is rooted in the challenges and demands of modern business environments, where effective communication, collaboration, and information management are crucial. Traditional methods of documenting and summarizing business meetings, such as manually taking notes or relying on memory, can be time-consuming, error-prone, and inconsistent. For instance, US11176949B2 suggests a concept for Devices, Systems, and Methods for Differentially Weighing Participants in a Meeting and for Automatically Generating Meeting Summaries. It is possible to generate or receive a transcript of the meeting,
3 | P a g e
and then analyze meeting data and meta-data using textual analysis and Natural Language Processing (NLP). Each participant is assigned a position or title, and depending on that job or title, a distinct weight is given to each participant's statements. The weight may also be adjusted based on the relationship between a participant's role and the topic of an utterance or by taking into account each participant's level of activity or passivity. A high-ranking participant may be given an optional predominant weight for all issues or just a select group of topics. A list of authorized choices, approved action items, rejected ideas, and other insights are automatically generated by the system.
Similar systems, techniques, and non-transitory computer readable storage media are described in US11689379B2, which are used to create meeting insights using media data and device input data. For instance, the disclosed system uses analysis of media data, such as audio or video data, and inputs to client devices connected to a conference to identify the parts of the meeting (e.g., the media data) that are pertinent to a user. The system generates an electronic message with content pertaining to the relevant component of the meeting after identifying the relevant part of the meeting. The user's client device receives the electronic message next from the system. For instance, the system generates meeting summaries, meeting highlights, or action items linked to the media data to present to the user's client device in one or more embodiments. In one or more embodiments, the system additionally trains a machine-learning model for use with subsequent meetings using the summary, highlights, or action items.
A system for providing an abstractive summary of a source textual document is disclosed in US10909157B2. The system consists of a fusion layer, a decoder, and an encoder. The encoder has the ability to create an encoding for the source text file. A linguistic model and a contextual model each make up the decoder. The contextual model can use the encoding to extract words from the source textual document. Based on pre-training with a training dataset, the language model is capable of creating vectors that paraphrase the underlying textual document. From the retrieved words and the generated vectors for paraphrasing, the fusion layer is able to produce the abstractive summary of the source textual document. In some implementations, the method makes use of a novelty meter to promote the creation of original terms for the abstractive summary.
4 | P a g e
The technique described in CN108804495B for automatic text summarization based on increased semantics includes the following steps: pre-processing entails organizing a text from high to low based on word frequency data and converting words to identification (id); employing a single-layer bidirectional LSTM to encode the input sequence and extract text information features; obtaining a hidden layer state by decoding the text semantic vector encoded using a single-layer unidirectional LSTM; generating a context vector, and then sifting between the input sequence and the current output to find the most pertinent information; Additionally, loss computation is done during the training step based on how semantically similar the fusion-generated abstract and the source text are improved. The invention uses the LSTM deep learning model to represent the text, integrates the semantic relationship between the context and the source text, improves the semantic relationship between the abstract and the source text, and generates an abstract that may be more appropriate for the text's theme. It also has a broad range of potential applications.
A system and method for producing summary details is given in US20200175050A1. The procedure calls for retrieving user-related event data. The event data is split into logical groups of similar user actions and attributes are derived from the event data. Each logical group's key details and related characteristics are recognized, and the user is assessed as a result. The user is shown the ranking groups of related activities that make sense. In response, a summary of the user's and logical groups' interactions is produced.
As was already noted, our invention's disclosures emphasize incorporating novel features and methods that set it apart from similar inventions that already deal with summary production. On the other hand, the concept of creating this model was advanced in order to be inspired by these discoveries while also bridging the gaps. The main components of the current invention, such as conceptual comprehension and multi-model integration, are the focus. a system that integrates sophisticated contextual awareness and goes beyond simple keyword extraction. This could entail determining speaker roles, assessing the conversation's flow, and picking up on subtleties that are important for precise descriptions. Combining text with audio and video to create a summary that is more thorough. This could entail writing summaries that incorporate both spoken and visual information, transcribing spoken content, and examining nonverbal signs. Integrate sentiment analysis to determine how emotionally charged the conversations in the meeting were. By giving important context for decision-making, this can
5 | P a g e
aid in understanding the participants' overall attitude and sentiment. By the promising solution that offers both businesses and people significant value.
Summary of the Invention
Business meeting summarizing is the process of presenting a succinct and fluid summary of a business meeting while maintaining the important information content and overall meaning. The purpose of using business summarization in business meetings is to help us summarize a recorded meeting while preserving important details and making sure the meeting's summary has the appropriate context and significance. Look into several business summary techniques. The summaries that were produced can be compared. Choose the most effective summary parameters (such as k in an extractive business meeting summary). Determine or, if necessary, alter an algorithm so that it may be scaled to multiple languages. Additionally, differentiate between the numerous ways that business meeting summaries are applied.In the proposed effort, we first transform recorded meetings, interviews, speeches, lectures, and other audio streams into text documents. The speaker text is then broken up into pieces. Using the suggested work, meeting minutes are automatically produced. Additionally, we concentrated on improving the voice-to-text conversion of a particular recorded audio file using the Rev-AI Speech-to-Text API while ensuring that the summarised text that has been converted from speech to text provides a specific and accurate meaning that is not only understandable to everyone but also covers all of the recorded file's key points.
Brief Description of Drawings
The invention will be described in detail with reference to the exemplary embodiments shown in the figure wherein:
Figure-1: Flowgorithm representing the methodology.
Figure-2: Flowgorithm representing speech to text conversion.
Figure-3: Extractive Summary Flow Chart.
Figure-4: Model Work Flow.
Detailed Description of the Invention
6 | P a g e
Utilizing cutting-edge machine learning techniques, meeting summarizing with artificial intelligence entails automatically condensing the important talks, conclusions, and outcomes of a meeting into a brief summary. AI can analyze spoken or written text to identify key points, produce concise summaries, and highlight action items, which helps with better decision-making and communication while also saving time.
The initial stage entails the attainment of research objectives as well as the gathering of essential meeting transcripts for research purposes. The information came from a corpus of roughly 75 meetings with a total running time of about 72 hours. An average of six people attended each meeting, producing transcripts with over 1000 lines that may be analyzed and condensed. The annotations or transcripts were properly prepared and stored using the NITE XML toolkit. There are several timestamps, tags, participant information, notes, and channels included in these meeting logs. The datasets are organized by signals, media, or audio and compressed in zip format. They must be properly processed before moving on because they have background noise and crosstalk, which requires conversion from.mrt to.txt format. The succeeding steps require a thorough examination of the meeting's setting and format, similar to a preliminary evaluation of the transcripts to ascertain the general scope of the content. To ensure the accuracy of the generated summaries, the next phase's pre-processing procedures or future actions are informed by this. Determining whether the meeting materials consist of one or more documents depends on the topics presented and the number of attendees after the meeting's domain has been established.
Pre-processing turns out to be a crucial step after data assembly and before using any summarizing approaches. The summarized meeting transcripts could contain jargon or background noise, leading to unsatisfactory summaries. As a result, the data needs to be pre-processed or refined before being subjected to direct analysis. The objective is to improve the quality of summaries generated from the dataset by removing superfluous words, symbols, special characters, and unfinished sentences as well as fixing errors in the transcript's words and phrases. As a result of the data being unalterably recorded, a sizeable amount of it contains background noise and crosstalk, which are dealt with in the pre-processing stage. Based on participant input and stop words, the meeting transcripts are processed and cleaned up. It was essential to consolidate the transcripts in order to fit the content into fewer lines given their
7 | P a g e
length and multi-file structure. 60 transcripts, each with 100 lines, were produced as a consequence of this process and put into a special folder for the following steps.
The dataset must be converted into a vector form before any operations to extract succinct summaries can be carried out. Text vectorization is a key component of Natural Language Processing (NLP) technology. In order to evaluate them, words or sentences are represented in numerous ways during this procedure. Word recognition and processing are aided by text vectorization. For data transformation, machine learning tools like scikit-learn are used. The most important terms in the transcript are identified using a method called TF-IDF (Term Frequency-Inverse Document Frequency), and scores are given depending on how significant those terms are in the context. Relevant features are extracted from the polished text and converted into vectors. These vectors are the basis on which NLP algorithms decide which key sentences should be included in the summary. In this methodology, the TF-IDF technique, which is used frequently in data extraction and summarization procedures, is combined with word frequency properties.
The pre-processed transcript is fed into the extractive summarization model as part of the Feature Extraction procedure. This model recognizes key elements and turns them into vectors, which NLP algorithms use to select only the most important sentences to include in the summary. This method combines the word frequency attribute with the term frequency-inverse document frequency (TF-IDF) approach, which is a standard technique in information retrieval and summarization. Based on the TF-IDF scores of the words used in those sentences, this technique calculates sentence scores. The term "term frequency" (TF) refers to how frequently a certain word appears in a transcript. The overall word count in the text is used to standardize the TF when dealing with meeting transcripts of various durations. The total number of transcripts is then divided by the number of transcripts that contain a certain word to produce the Inverse Document Frequency (IDF).
Two methods of text summary are frequently used: extractive and abstractive summarizations. While the abstractive technique is emphasized at this time, some aspects of the extractive method are also included. Natural Language Processing (NLP) is integrated with deep learning methods such as long short-term memory (LSTM) and sequence-to-sequence. The TextRank algorithm, a graph-based NLP ranking method inspired by Google's PageRank algorithm, is used in our strategy. For machine learning and NLP tasks, we built models using
8 | P a g e
LSTM and RNN, utilizing programs like scikit-learn, TensorFlow, deep neural networks, as well as different Python libraries. In the NLP field, RNN stands out as a common neural network design that works quickly and accurately to create language models and tackle speech recognition problems. RNNs can only learn from recent data, however, because of their understanding of context and immediate dependencies. LSTM gets beyond this restriction, making it a unique RNN variation that excels at gathering contextual data.
In this step, the TextRank algorithm—originally developed from Google's PageRank algorithm—transforms the vectorized representation of phrases into a visual representation. The foundation of this algorithm is a graph-based ranking method. While the nodes themselves in the graph represent individual sentences, the connections between the nodes reflect the similarity scores between sentences. An extractive form is created by choosing and combining the best-ranking phrases, which is based on a predetermined threshold value.
The extractive approach is used to create summaries, which are then integrated with the original text. The result is a CSV file that has two columns with the labels "text" and "summary." The textual content of the transcripts and their related extracted summaries are contained in these columns. This merged dataset is then used as input for the abstractive technique, allowing for the creation of a summary using that methodology.
The summary created by the extractive process in the first stage is used as input in the second step of the abstractive technique. The original text and this summary are both supplied in CSV format. Neural networks are used to implement a deep learning method in this stage. Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs) stand out among the frequently used neural network architectures.
CNNs are especially effective for applications like facial recognition and picture classification, where the input length is fixed and the results depend solely on the current input at each level. Both the encoder and decoder portions of the networks in this design use RNNs. RNNs use data from recent inputs and previous outputs to anticipate the outcome of the current state. RNNs, on the other hand, show poor effectiveness for incredibly brief periods. Long Short-Term Memory (LSTM), on the other hand, is an improved RNN variation that successfully handles long-term temporal dependencies.
9 | P a g e
The procedure also incorporates an Attention mechanism. By focusing on certain input segments, this method enables the anticipation of output for lengthy sequences. Therefore, an RNN LSTM Sequence-to-Sequence Decoder-Encoder combined with an attention mechanism is the best configuration for the second stage of the text summarization approach. The encoders in this configuration go over the entire input sequence and encode it into a fixed-size context vector. The decoder then decodes this data to produce the final output, an abstractive summary.
The created model now has to be trained and validated. To accomplish this, a 70:30 split of the input text and summary dataset in CSV format is used for testing and training. Several hyperparameters, including the optimizer, epochs, batch size, loss function, learning rate, and embedding dimension, are adjusted during the model's training. Each sample's testing and training accuracy as well as loss are calculated during the training procedure. The model's training process is thus more clearly shown using these measures, which show decreased loss and improved accuracy.
According to the conclusions drawn from the literature analysis, there is no predetermined set of scores that have significance beyond their values. Evaluation instead makes use of either a ROUGE score or a human evaluation. According to the aims, the generated summaries must be succinct, clear, understandable, and coherent while retaining the fundamental ideas of the material.
In conclusion, the process can be roughly split into two parts. The first section describes the steps that were performed to process the input transcripts, which eventually resulted in the creation of potential summary outputs. The next part explains how to use ROUGE metrics to assess the summaries produced by both the abstractive strategy, which includes LSTM, RNN, and an attention mechanism, and the extractive approach, which includes TextRank. ADVANTAGES OF THE PROPOSED MODEL
Business Meeting Summary Generation has a number of benefits that can help people and companies in many different ways. Due to the length of some business meetings, attendees may not have enough time to listen to the complete recording or text. Hours of
10 | P a g
e
conversation can be condensed into brief summaries using NLP-powered summary generation, enabling attendees to rapidly understand the important ideas and conclusions of the meeting.
Reading a summary takes far less time than listening to or reading all of the meeting's materials. Participants can keep informed thanks to this effectiveness without having to devote too much time to meeting-related tasks. Individuals who might not have been able to attend the meeting owing to schedule issues or other reasons can nevertheless receive the meeting's material thanks to summaries. By being inclusive, everyone can continue to be informed and involved.
Complex discussions can be condensed into concise, well-focused summaries using NLP algorithms. This helps to prevent misconceptions and makes sure that important points are correctly noted. The quality and content of automated NLP-generated summaries are consistent. There is no chance of overlooking crucial information owing to bias or human error, unlike manual note-taking.
Summaries produced by NLP offer a dispassionate view of the meeting's substance. This might be helpful for going back and examining choices, assessing tactics, and spotting trends over time. The purpose of summaries is to provide quick access to information for upcoming conversations, action items, and project planning. Members of the team can review summaries to review prior discussions and decisions.
Summaries may be produced for international teams that interact in a variety of languages more easily since NLP models can be trained to deal with numerous languages. NLP-powered solutions are scalable for firms with regular meetings or big teams since they can handle a huge volume of sessions and produce summaries rapidly. In the long run, the time and effort saved by automating summary production might lead to cost savings, even though early setup and integration of NLP systems may need investment. , Claims:The scope of the invention is defined by the following claims:
Claims: 1. The System/Method for Business Meeting Summary Generation (BMSG) comprising:
a) The need for human note-taking and summary is eliminated by automating the generation of accurate and succinct summaries for business meetings using NLP techniques. b) In order to ensure that the generated summaries are focused on the crucial information, the most crucial and pertinent aspects are extracted from dialogues. The initiative asserts to eliminate the possibility of bias or human error in note-taking while providing reliable summaries that accurately and consistently record important choices, actions, and discussions.
c) The automated summaries will open up conference content to a wider audience, including people who couldn't physically attend the meeting. With the initiative, participants will be able to grasp debates without having their own personal prejudices or interpretations cloud their understanding.
2. According to the claim 1, the created summaries will make it easier for teams to collaborate since they will make it possible for them to rapidly comprehend the key points of other teams' discussions.
3. According to the claim 1, the summaries produced will be used as quick references for going over previous decisions, keeping track of action items, and planning upcoming projects.