I've had some good exposure to speech recognition and understanding technologies, including productive interactions with some of the luminaries in the field. I am incredibly impressed with what they have achieved, and I suspect that the field of robotics will continue to spin off ideas and applications that have relevance to theoretical linguistics.
Text processing research has already yielded benefits for theoreticians. Towards the end of his life, Chuck Fillmore pointed out that one of the AI researchers who had the greatest effect on his thinking was Roger Schank, a very vocal critic of theoretical linguists (albeit not as much of generative semanticists). Since I had shifted to the field of Natural Language Processing and AI research after 1990, I began to run into Fillmore at Association for Computational Linguistics conferences. CL had come to have a great interest for him, especially as the major developer of
FrameNet at Berkeley.
You can see what impressed Fillmore so much in the early work of Schank, Abelson, and others on conversational slot-filler programs. Schank basically showed that a conversation could not be interpreted coherently without information that existed outside the overt linguistic signal. For example, he would describe a conversation about a visit to a restaurant in terms of shared "world knowledge" in the background, which Schank implemented as a descriptive list of events, activities, and roles that the conversation would be about. So he would look at the inferences that would license understanding of narratives like:
1. John was hungry, so he went to the diner.
2. The food was so bad that he could not finish.
3. He left without leaving a tip.
Schank represented the "world knowledge" in terms of a generic script that would allow a computer program to instantiate information from the text and answer questions about the text that depended on the unspoken information in the script. For example, you can answer the question "Do you think the waiter was happy?", because the script would contain information that waiters served food, received tips, and were happy to receive tips. The script might also contain information about what tips are and the motivation for leaving them or not leaving them. So the coherence of those three sentences as a narrative really depends on information that is not in the linguistic signal--for example, no reference whatsoever to a "waiter", even though one could ask questions about the probable state of mind of a waiter from reading that conversation.
This kind of observation really inspired Fillmore. So he and his colleagues at Berkeley decided to build a repository of more sophisticated "scripts" in terms of recursively-structured networks of "frames" that one could map to words in a lexicon. He could then describe a "family" of words in terms of an explicit representation of the concepts that tied the words together. As a lexicologist, he delighted in the power that the idea gave him. There are plenty of limitations to his approach, but the result is a set of information structures that can actually be used to extract useful information from text. So FrameNet became one of several useful tools that NLP researchers can use to perform useful information extraction from text.