International Business Machines Corporation

Hewlett-packard Development Company, L.p.

Dealigence Inc.

David Konopnicki

Haggai Roitman

Guy Feigenblat

Michal Shmueli-Scheuer

Odellia Boni

Doron Cohen

Laurent David Hasson

Jonathan Herzig

Tommy Sandbank

David Andrew Brooks

Yosi Mass

Tal Drory

Igor Belakovskiy

Lin Sun

Oded Shmueli

Thomas R Haynes

David Carmel

Oded Margalit

Noam Slonim

Ranit Aharonov

Boaz Carmeli

Nitzan Peleg

Artem Spector

Priscilla Santos Moraes

Dafna Sheinwald

Amir Bar-Or

Ilya Shnayderman

Ella Rabinovich

Benjamin Sznajder

Gilad Barkai

Guy Lev

Lior Leiba

Yafit Sami

Yehoshua Sagiv

Or Rivlin

Achiya Jerbi

Gadi Aleksandrowicz

Edward Moroshko

Ora Peled Nakash

Mordechai Taitelman

Hen Shkedi

Clarification-question selection, including: Receiving a search conversation that includes utterances by a user and by a conversational search system. Retrieving, from a solution documents database, text passages that are relevant to the search conversation. Retrieving, from a clarification questions database, for each of the text passages, candidate clarification questions that are relevant to both the respective text passage and the search conversation. Ranking the candidate clarification questions according to (a) a first deep learning model that outputs, for each of the candidate clarification questions, a score denoting a strength of association between the respective candidate clarification question and the search conversation, and/or (b) a second deep learning model that outputs, for each of the candidate clarification questions, a score denoting a strength of association between the respective candidate clarification question, the search conversation, and the text passage that was retrieved for the respective candidate clarification question.

Clarification questions selection for conversational search, trained from human-to-human conversations and solution documents

An example system includes a processor to receive concepts extracted from a result set corresponding to a query and result associations for each extracted concept. The processor is to build a graph based on the extracted concepts, wherein the graph comprises a number of nodes representing the extracted concepts and weighted edges representing similarity between concepts extracted from shared results. The processor is to partition the graph into subgraphs with vertices corresponding to candidate facets for vertices having higher sums of weighted edges. The processor is to rank the candidate facets. The processor is to select higher ranked candidate facets to use as facets. The processor is to output facets with a result set in response to the query.

Dynamically generating facets using graph partitioning

An example system includes a processor to receive a query. The processor can retrieve ranked candidates from an index based on the query. The processor can re-rank the ranked candidates using a Bidirectional Encoder Representations from Transformers (BERT) query-question (Q-q) model trained to match queries to questions of a frequently asked question (FAQ) dataset, wherein the BERT Q-q model is fine-tuned using paraphrases generated for the questions in the FAQ dataset. The processor can return the re-ranked candidates in response to the query.

Frequently asked questions and document retrieval using bidirectional encoder representations from transformers (BERT) model trained on generated paraphrases

Embodiments may provide techniques to generate training data for summarization of complex documents, such as scientific papers, articles, etc., that are scalable to provide large scale training data. For example, in an embodiment, a method may be implemented in a computer system and may comprise collecting a plurality of video and audio recordings of presentations of documents, collecting a plurality of documents corresponding to the video and audio recordings, converting the plurality of video and audio recordings of presentations of documents into transcripts of the plurality of presentations, generating a summary of each document by selecting a plurality of sentences from each document using the transcript of the that document, generating a dataset comprising a plurality of the generated summaries, and training a machine learning model using the generated dataset.

Automatic generation of training data for scientific paper summarization using videos

Automated keyphrase extraction from a digital text document. A pool of candidate keyphrases of the digital text document is created. A cross-entropy method is then employed to compute a set of output keyphrases out of the pool of candidate keyphrases, by iteratively optimizing an objective function that is configured to cause the set of output keyphrases to be descriptive of one or more main topics discussed in the digital text document. The set of output keyphrases may be used for at least one of: text summarization, text categorization, opinion mining, and document indexing.

Automatic keyphrase extraction from text using the cross-entropy method

A method, computer system, and computer program product for generating a multi-document summary is provided. The embodiment may include receiving a query statement, one or more documents, one or more summary constraints, and quality goals. The embodiment may include identifying one or more keywords within the query statement. The embodiment may include performing a sentence selection from the one or more documents based on the one or more identified keywords. The embodiment may include generating a plurality of candidate summaries of the one or more documents based on the performed sentence selection, the goals, and a cross entropy method. The embodiment may include calculating a quality score for each of the plurality of generated candidate summaries using a plurality of quality features. The embodiment may include selecting a candidate summary from the plurality of generated candidate summaries with the highest calculated quality score that also satisfies a quality score threshold.

Extractive query-focused multi-document summarization

A method, computer system, and a computer program product for optimizing a plurality of personality traits of a virtual agent based on a predicted customer satisfaction value is provided. The present invention may include identifying a customer. The present invention may also include retrieving a plurality of data associated with the customer. The present invention may then include analyzing the received plurality of data using a customer satisfaction prediction model. The present invention may further include generating a plurality of analyzed data from the customer satisfaction prediction model based on the analyzed plurality of data. The present invention may also include generating a plurality of personality traits for a virtual agent from the generated plurality of analyzed data.

Optimizing personality traits of virtual agents

A computerized method comprising receiving, from a question answering system, a minimal answer value to a query submitted by a user. Also received are electronic documents based on the minimal answer value, and a document score value, associated with the query, for each of the electronic documents. The method comprises extracting entities and attributes from electronic documents, and for each computing one or more associated score value, and aggregating the document score value with the associated score values. The method comprises selecting some of entities and attributes based on the respective aggregated score value, thereby producing selected associated elements. The method comprises generating, using a computerized natural language (NL) generating system, a comprehensive NL answer, wherein the generating is based on the minimal answer value and the selected associated elements, and sending the comprehensive NL answer for presentation to the user.

Generating natural language answers automatically

Embodiments of the present systems and methods may provide techniques by which bots may be analyzed using improved representations of bot structure and a means of assessing conversational quality that may provide improved efficiency. For example a method may comprise training, at a computer system comprising a processor, memory accessible by the processor, and computer program instructions stored in the memory and executable by the processor, a neural network model to learn representations that capture characteristics of the graphs of chatbots, wherein the captured characteristics include at least a content-based representation based on user utterances that are relevant to the nodes and based on the chatbot response for the nodes.

Neural representation of automated conversational agents (chatbots)

Method and apparatus for training and using a classifier for words. Embodiments include receiving a first plurality of sentences comprising a first word that is associated with a class and a second plurality of sentences comprising a second word that is not associated with the class. Embodiments include training a classifier using positive training data for the class that is based on the first plurality of sentences and negative training data for the class that is based on the second plurality of sentences. Embodiments include determining a measure of correlation between a third word and the class by using a sentence comprising the third word as an input to the classifier. Embodiments include using the measure of correlation to perform an action selected from the following list: selecting content to provide to a user; determining an automatic chat response; or filtering a set of content.

Classification engine for learning properties of words and multi-word expressions

A method comprising receiving digital documents, a query statement, and a summary length constraint; identifying, for each of said digital documents, a sentence subset, based, at least in part, on said query statement, a modified version of said summary length constraint, and a first set of quality objectives, generating, for each of said sentence subsets, a random forest representation; iteratively (i) sampling, from each of said random forest representations, a plurality of tokens to create a corresponding candidate document summary, based, at least in part, on weights assigned to each of said tokens, (ii) assigning a quality ranking to said candidate document summary, based, at least in part, on said first set of quality objectives and a second set of quality objectives, and (iii) adjusting said weights, based, at least in part, on said quality rankings; and outputting a highest ranking said candidate document as a compressed summary.

Context-aware sentence compression

Embodiments may provide automated summarization of documents, such as scientific documents by using a prior distribution on logical sections learnt from a corpus of human authored summaries. For example, a method of document summarization may comprise receiving, at the computer system, a document and segmenting the document into a plurality of sentences, identifying, at the computer system, sections in the document and aligning each sentence in the document to a section logical role, and summarizing, at the computer system, the document using a probability distribution.

Unsupervised attention based scientific document summarization

A system for generating a summary of a text document is disclosed. In some examples, the system includes a processor configured to generate an initial summary of an original document. The initial summary includes a selection of extracted sentences copied from the original document. For each extracted sentence of the initial summary, the processor processes the extracted sentence to generate an abstracted sentence, and generates vector representations of the extracted sentence, the abstracted sentence, the original document, and the current summary. The vector representations are then input to a decision network to compute an editing decision. The editing decision is selected from a group of possible decisions that includes a decision to add the extracted sentence and a decision to add the abstracted sentence. The processor also updates the current summary based on the editing decision.

Natural language processing techniques for generating a document summary

A system comprising at least one hardware processor; and a non-transitory computer-readable storage medium having stored thereon program instructions executable to receive, as input, one or more digital documents, a query statement, and a summary length constraint, automatically generate, for each of said one or more digital documents, an initial summary based, at least in part, on a first sentence selection which satisfies said query statement, a modified said summary length constraint, and a first summary quality goal, automatically extract, from each of said initial summaries, one or more associated feedback metrics, and automatically generate, for each of said one or more digital documents, a final summary based, at least in part, on: (i) a second sentence selection which satisfies said query statement, said summary length constraint, and a second summary quality goal, and (ii) at least one of said associated feedback metrics.

Query-based extractive summarization

For each intent associated with a feature in a discordant conversation, one or more preceding discordant user utterances and one or more following discordant user utterances are collected. A discordant distribution over terms of the one or more preceding discordant user utterances and the one or more following discordant user utterances is created. For each intent associated with a feature in a non-discordant conversation, one or more preceding non-discordant user utterances and one or more following non-discordant user utterances are collected. A non-discordant distribution over terms of the one or more preceding non-discordant user utterances and the one or more following non-discordant user utterances is created. The discordant and non-discordant distributions are compared and the top-k terms that are most specific to user utterances associated with the corresponding discordance feature using Kullback-Leibler divergence are determined. An intent dominancy score is generated to provide an analytical view of the discordant conversation.

Growing community of inventors

Haifa, Israel

David Konopnicki