Avis sur HumanFirst
Once I used it I realized it was what I'd been dreaming of
Commentaires : The UI challenges are so insignificant in comparison to what HF offers. The HF team are intelligent and kind. The tool works and we are getting value from it.
The tedious nuance of clustering and comparing outputs is abstracted away from the engineer who can work more efficiently. It lets you build agents from your chat data at scale revealing customer intents and the relation to your existing model.
The UI needs some work. It's a little confusing to know which model is active at any given time. Some functionalities are not intuitive.
Alternatives envisagées précédemment :
Commentaires : Initially, we used HumanFirst for primarily for intent training, and have trained and continue to refine several classifiers. However, we also recently switched to using HumanFirst as our real-time NLU for our chatbots by integrating it into our bot building tool, and have seen a ~10-20% rise in intent classification since (without any additional retraining).
HumanFirst lets you upload a lot of data then sort through it methodically with the aid of several clustering, filtering and suggestive features. It's made our data labelling process much more efficient, and has guided us while we continue to iteratively improve and extend our models. With HumanFirst, you can quickly put together an oversight of the data and notice groups that might have otherwise been missed. The UI is clean and intuitive though constantly evolving as new features are made available. The HumanFirst team are genuinely receptive to feedback and respond to our suggestions of what we'd like to see in future, as well as helping all team members get familiar with the product and suggesting ways of using HumanFirst tools in use cases that we hadn't considered of before.
Occasionally there is the odd bug but it's often quickly fixed after we flag it to the HumanFirst team. There are some areas that would be really useful to us that could be developed further (like entity training), but I believe the team is working on this.
Commentaires : After we boot-strapped the AI by assigning intents to almost all of the articles and added the SEO tags to the articles, we saw an immediate increase in CTR. I know this can be further improved through more gap analysis and intent tagging. I was also able to identify several gaps and write new articles for them, with dozens more on the way. HF's support has always been very helpful with any challenges I have and to try to improve the tool.
I love the way that HF can help me identify what customers are searching for by grouping intents. The searches are captured in Zendesk but without a way to organize that information it has little value. Organizing these searches by intent in HF allows me to improve SEO and CTR by adding the intents to the existing articles, and also helps with gap analysis for topics we haven't covered yet.
I sometimes have challenges with the AI misidentifying topics and clusters but that may just be due to the nature of my business. We work across multiple media and have lots of equipment so the AI sometimes has difficulty identifying what search goes with which product unless the customer specifies it. I don't know if there's a solution to this problem since customers will inevitably make vague searches, but it makes for a lot of dead wood in the searches since I don't want to assign a vague search and mis-train the AI if I incorrectly assume the customer's intent. It may help if I could assign the same intent to multiple articles. For example, when a customer searches for "parental control" I want to make sure they get results for parental control on their Wi-Fi network, and parental control on each of our cable boxes.
The tool that turned conversation designers, into NLU designers
Commentaires : The overall experience has been overwhelmingly positive. From a toolset perspective: We can build NLU models extremely fast, disambiguate between intents, and use the tool within our BAU process for continuous improvement. On average, the models that we've used Human First, have about a 10-15% higher F1 score than the previous. From working with them: Human First feels more like a partner. They're deeply invested in our success, which shows with the fast communication, weekly syncs and input that we've had into the product itself.
The ability to cluster semantically similar text has been a game changer for us. Rather than trudging through excel spreadsheets looking for similar utterances, we could load tens of thousands of data points into Human First and completely rebuild a failing bots NLU model within weeks (rather than months). The disambiguation feature is also brilliant. Using this, we can better understand why conflicts are arising and fix them. To add to this, now, with the CX integration, we can make these changes knowing they'll work within our DialogFlow bots. A tool like Human First has been like upgrading from MS Paint to Photoshop. Rather than living within spreadsheets, we can apply a more "design thinking" approach to developing our NLU models.
Being new-ish software, there are sometimes some UI quirks and strange error messages. Due to the size of some of our bots, we have, once or twice, uploaded too much data into the tool, which resulted in some projects having trouble rendering it. However, this has never been a significant issue for us as the development team generally fixes problems within 24 hours.
No-code NLU for bot at scale
Commentaires : HumanFirst helped the team move to the next level. They listened to our needs and adapted the product according to our requirements. They supported the team members so that they were able to use the product optimally according to our use cases. I recommend HumanFirst for their product and for their team. They are good collaborators and business partners.
What I like the most is the bottom-up labeling workflow. They offer a smart way to perform a task that was a bit mind-numbing when done manually. Their approach allows you to really be data-driven and to process large datasets by focusing on what really improves the model instead of labeling everything.
For our business, integration would be easier if the product was available in a cloud platform marketplace.