Mitsubishi Electric Develops Dialogue-summarizing AI Based on Knowledge Processing

Mitsubishi Electric Corporation announced today that it has developed a dialogue-summarizing technology based on its AI technology Maisart ®* that automatically generates accurate written summaries of spoken conversation, thereby enabling call-center staffers to roughly halve the time they need to prepare final reports after calls, as determined in preliminary tests conducted at a Mitsubishi Electric call center.

*M itsubishi Electric’s AI creates the S tate-of-the- ART in technology

Fig. 1: Example application at call center

Key Features

1. Automatically generates written records and halves final report preparation time

Recordings of call-center dialogues prepared with conventional speech-recognition technology typically contain extraneous utterances, broken speech, etc., making it difficult for the software to generate useful written records of the dialogues.

Mitsubishi Electric’s new technology (Fig. 2), however, learns dialogue contexts to accurately determine meanings, including for colloquial and semantical expressions. It then uses data from past reports to correct broken speech, grammatical errors, word choices, etc. to generate complete natural sentences.

Next, it extracts the most similar and shortest sentences from past reports edited by staffers and then finally produces a summarized written record. Internal tests showed that Mitsubishi Electric’s new technology not only dramatically condenses Japanese text volume, about 90% of the text is usable compared to about 30% in the case of conventional technology, which enables call-center staffers to roughly halve the time they require to manually prepare their final reports.

2. Incrementally learns manual corrections to gradually improve correction accuracy

When people call to inquire about new products, etc., it is difficult to generate accurate records of these conversations automatically since conventional software cannot refer to call histories and past reports covering similar topics. To overcome this obstacle, Mitsubishi Electric's new AI technology steadily refreshes and updates its knowledge by referring to recently edited reports to learn how staffers manually convert broken syntax, colloquial expressions, homonyms, etc. into more natural language.

Comparison of new and conventional methods

	Dialogue processing method	Performance
New	Maisart learns call histories and past reports to create a written record, and then extracts the most similar and shortest sentences from manually produced past reports to generate a condensed record of summarized natural sentences	About 90% of generated record usable in final report
Conventional	AI learns how to convert spoken sentences into summarized sentences through end-to-end machine learning.	About 30% of generated record usable in final report

Future plans

Mitsubishi Electric will evaluate its new technology through demonstration tests at the company's own call center from March 2021, aiming eventually to use it to generate records of phone conversations with customers regarding product malfunctions, product inquiries, etc.

For the full text, please visit: www.MitsubishiElectric.com/news/

March 17, 2021